Journal of Pranjul Rathour · 286 articles
Writing, from the work.
Build logs from real systems, hackathon stories from both sides of the judging table, guides for GenAI engineering, and notes from mentoring 200+ students. Every claim traces back to something shipped, won or taught.
Pranjul Rathour · GenAI Engineer, Kanpur, India · open to GenAI roles, hackathon judging, mentoring and campus talks.
Also written as Pranjal Rathour, Pranjul Rathore, Pranjal Rathore, Pranjul Rathor, Pranjal Rathor, Pranjul Rathaur, Pranjul Rathur, Pranjull Rathour.
Latest · Mentoring & speaking
From mentee to mentor to judge: a natural progression, not a hierarchy
The path from being mentored, to mentoring others, to judging their work isn't a ladder to climb — it's a set of roles the same person can hold at once, at any stage.
01RAG & retrieval36 articles
- RAG & retrieval
Why my RAG platform says "I don't know" — building RAG.NextUpgrad
Inside RAG.NextUpgrad: hybrid dense + BM25 retrieval, cross-encoder reranking, and the confidence gate that refuses to answer on weak evidence.
2 min readRead - RAG & retrieval
Duplicate and near-duplicate documents in a RAG index: the quiet quality killer
The same policy document uploaded three times with minor edits crowds out other results and confuses retrieval. How to detect and handle near-duplicates.
1 min readRead - RAG & retrieval
RAG over a codebase: why retrieving code needs a different chunking strategy
Chunking code by token count the same way you'd chunk prose breaks function boundaries and destroys context. A code-aware retrieval pipeline needs a different approach.
1 min readRead - RAG & retrieval
How much retrieved text should actually go in the prompt? A context-budget framework
More context isn't always better — past a point, extra chunks dilute attention and cost more without improving the answer. A practical way to pick the cutoff.
1 min readRead - RAG & retrieval
RAG chatbot vs RAG-powered search: same technique, different product
A conversational chatbot and a search bar built on the same RAG pipeline need different context handling, different UX, and different failure-mode tolerance.
1 min readRead - RAG & retrieval
Evaluating retrieval quality when you have no labelled query-document pairs
Proper retrieval evaluation needs labelled relevance judgements you probably don't have yet. Practical ways to get a useful signal before that dataset exists.
1 min readRead - RAG & retrieval
RAG over mixed Hindi-English documents: what breaks and what to check
Indian document sets often mix Hindi, English and transliterated text on the same page. Standard embedding models handle this unevenly — what to test before trusting them.
1 min readRead - RAG & retrieval
Multi-tenant RAG: making sure one customer never sees another customer's documents
A shared vector index serving multiple customers needs strict isolation, not just a metadata filter that a bug could silently bypass. What actually guarantees separation.
1 min readRead - RAG & retrieval
The RAG security hole nobody checks: prompt injection hiding inside retrieved documents
A malicious instruction embedded in a document your RAG system retrieves can hijack the model's behaviour. What this attack looks like and how to guard against it.
1 min readRead - RAG & retrieval
Incremental indexing: keeping a RAG index current without re-embedding everything
Documents change. Re-embedding an entire corpus on every update is wasteful and slow. A practical approach to updating only what changed.
1 min readRead - RAG & retrieval
RAG vs a search engine: what users actually expect, and why the gap causes complaints
Users compare a RAG chatbot to Google, not to a research assistant. That mismatch in expectation causes most of the 'it's wrong' complaints a RAG app gets.
1 min readRead - RAG & retrieval
What a confidence gate is, and why every RAG app needs one
A confidence gate stops an LLM from answering when retrieval comes back weak. Here's how to design the threshold and what it costs to skip it.
1 min readRead - RAG & retrieval
Hybrid retrieval explained: dense search, BM25, and Reciprocal Rank Fusion
Why combining vector search with keyword search beats either alone, and how Reciprocal Rank Fusion merges the two without score normalisation headaches.
1 min readRead - RAG & retrieval
RAG systems: a complete beginner's guide (2026)
What retrieval-augmented generation actually is, why it exists, and the components a real RAG system needs beyond a vector database.
1 min readRead - RAG & retrieval
RAG vs fine-tuning: which one your problem actually needs
A decision guide from someone who has shipped both: when retrieval solves it, when fine-tuning solves it, when you need both, and the questions to ask before spending a GPU hour.
2 min readRead - RAG & retrieval
Query rewriting for RAG: expansion, decomposition and HyDE, explained simply
Users ask messy questions. Query rewriting turns them into retrievable ones: synonym expansion, splitting multi-part questions, hypothetical document embeddings, and how to tell if any of it helps.
1 min readRead - RAG & retrieval
Seven ways RAG fails in production, and the fix for each
The failure modes that show up after the demo: stale indexes, wrong-document leakage, boundary splits, confident refusals, prompt injection via documents, cost blow-ups and silent provider drift.
2 min readRead - RAG & retrieval
A RAG project for your college: notes, syllabus and PYQs you can actually query
A complete student project plan: build a retrieval-augmented assistant over course notes, syllabus PDFs and previous-year questions, with the architecture, data work and demo that make it portfolio-worthy.
2 min readRead - RAG & retrieval
Multi-provider LLM fallback: staying up when one API goes down
How to design automatic fallback across LLM providers: provider abstraction, health checks, prompt portability, cost-aware routing and the failure cases that a retry loop cannot fix.
2 min readRead - RAG & retrieval
Streaming LLM answers with FastAPI and Server-Sent Events
How to stream tokens from an LLM to a browser with FastAPI and SSE: the endpoint shape, the client, error handling mid-stream, and the pitfalls with proxies and timeouts.
2 min readRead - RAG & retrieval
Citations in RAG answers: how to make every sentence traceable to a source
Design patterns for citation-grounded answers: chunk identifiers, inline markers, page references, and how to verify that a citation actually supports the claim it is attached to.
2 min readRead - RAG & retrieval
Choosing an embedding model for RAG (and when to switch)
How embedding models differ, what to test before committing, why multilingual matters in India, and how to swap models later without re-architecting the pipeline.
1 min readRead - RAG & retrieval
FAISS vs Qdrant vs pgvector: choosing a vector store for a student or startup RAG project
Three vector stores I have actually used, compared on setup, filtering, persistence, hosting cost and when each one is the right call for a RAG project.
2 min readRead - RAG & retrieval
How to evaluate a RAG system: recall, faithfulness and the questions that matter
A practical evaluation setup for retrieval-augmented generation: build a question set, measure retrieval recall, grade faithfulness and refusal accuracy, and track it on every change.
2 min readRead - RAG & retrieval
Reranking in RAG: why a cross-encoder second pass fixes most bad answers
Bi-encoders retrieve fast but rank loosely. A cross-encoder reranker reads query and passage together and reorders the top candidates — here is how, what it costs, and when to skip it.
2 min readRead - RAG & retrieval
Chunking strategies for RAG: sizes, overlap and structure-aware splits
How to split documents for retrieval-augmented generation: chunk sizes that retrieve precisely, overlap that preserves context, and structure-aware splitting for PDFs and code.
2 min readRead - RAG & retrieval
Self-hosted vs managed vector search: what actually changes at student-project scale
pgvector on a free Postgres instance vs a managed vector database — the real trade-offs for a project with under a million vectors, not an enterprise deployment.
1 min readRead - RAG & retrieval
RAG observability: the five things worth logging (and the ten that aren't)
Logging everything drowns the one signal that matters. A short list of what to actually capture from a production retrieval pipeline, and why.
1 min readRead - RAG & retrieval
The RAG question nobody plans for: 'what if it's just not in the documents?'
A confidence gate catches weak matches, but out-of-scope questions need a different check: does the retrieved context even address the question's topic at all?
1 min readRead - RAG & retrieval
RAG over a SQL database: why text-to-SQL is a different problem
Vector search doesn't help when the answer is 'run this aggregate query'. A practical pattern for letting an LLM answer questions against a real database safely.
1 min readRead - RAG & retrieval
Multi-hop questions break single-pass RAG. Here's the pattern that fixes it
'Who won the hackathon the year after X joined?' needs two retrieval steps, not one. A simple decompose-then-retrieve pattern for multi-hop questions.
1 min readRead - RAG & retrieval
Where a RAG request's latency actually goes (it isn't the LLM call)
Profiling a real RAG endpoint end to end: embedding, vector search, reranking, prompt assembly and generation each take a slice — and the slowest part surprises most people.
1 min readRead - RAG & retrieval
Semantic caching for RAG: when it saves money and when it lies to users
Caching by exact string match misses near-duplicate questions; semantic caching catches them, but a badly-tuned similarity threshold serves stale, wrong answers.
1 min readRead - RAG & retrieval
RAG for PDFs with tables and images: what breaks and how to fix it
Plain-text chunking destroys tables and ignores images. A practical pipeline for PDFs that actually contain the data your users are asking about.
1 min readRead - RAG & retrieval
Metadata filtering in RAG: the retrieval fix nobody mentions first
Before you touch embeddings or rerankers, filter by metadata — date, source, document type. It is the cheapest accuracy gain in a retrieval pipeline.
1 min readRead - RAG & retrieval
I ran the chunk-size experiment so you don't have to: what actually changed
Testing 128, 256, 512 and 1024-token chunks on the same retrieval pipeline, with the same eval set, to see which size actually moved recall and answer quality.
1 min readRead
02Fine-tuning33 articles
- Fine-tuning
Fine-tuning a 1.7B model at 3.2 GB VRAM — building FineTune Studio
How FineTune Studio runs QLoRA fine-tuning end to end in a browser, on hardware a student can actually afford, with honest base-vs-tuned evaluation.
1 min readRead - Fine-tuning
Evaluating fine-tuning ROI: was the project actually worth the compute and time?
After a fine-tune ships, it's worth asking honestly whether it beat the simpler alternative. A framework for answering that instead of assuming success.
1 min readRead - Fine-tuning
Fine-tuning a vision-language model: what's different from fine-tuning a text-only LLM
Multimodal fine-tuning adds an image encoder into the picture, literally. What changes in data format, compute needs and evaluation compared to text-only fine-tuning.
1 min readRead - Fine-tuning
Transfer learning vs fine-tuning: are these the same thing?
The terms get used almost interchangeably, but they're not quite the same idea. A clear distinction, and why it rarely matters practically but sometimes does.
1 min readRead - Fine-tuning
Estimating fine-tuning cost before you start, so there are no surprises
GPU-hours, dataset size and number of epochs multiply into a real number before you've spent anything. A back-of-envelope way to estimate it upfront.
1 min readRead - Fine-tuning
Continual fine-tuning: updating a model over time without starting from scratch
New data arrives every month. Retraining from the original base model every time is wasteful. What continual fine-tuning gets right, and its real risk.
1 min readRead - Fine-tuning
Fine-tuning for reliable structured output (JSON mode isn't always enough)
A model that occasionally breaks a JSON schema under prompting alone can be fine-tuned to be far more reliable. When this is worth doing over prompt engineering.
1 min readRead - Fine-tuning
Deduplicating a fine-tuning dataset before training (and why near-duplicates matter more than exact ones)
Exact duplicates are easy to catch. Near-duplicates — the same example paraphrased — quietly bias a fine-tune and are far more common than people expect.
1 min readRead - Fine-tuning
Fine-tuning an open-weight model vs a closed API's fine-tuning offering
Some providers offer hosted fine-tuning on their closed models. The trade-offs against fine-tuning an open-weight model yourself are real and worth knowing upfront.
1 min readRead - Fine-tuning
Reward modeling, explained simply: what it is and why most student projects don't need it
Reward models power RLHF-style training and are genuinely complex to build well. A plain-language explanation, and when a student project actually needs one.
1 min readRead - Fine-tuning
Fine-tuning a model for a specific tone of voice: what actually transfers
Getting a model to consistently sound like your brand or persona is one of the cleaner wins fine-tuning offers. What data actually teaches tone, and what doesn't.
1 min readRead - Fine-tuning
How to evaluate a fine-tuned model honestly (base vs tuned)
If you can't show the before and after on the same prompts, you didn't fine-tune — you spent GPU hours. Here's how to evaluate properly.
1 min readRead - Fine-tuning
QLoRA fine-tuning: a complete guide for students
What QLoRA actually does differently from full fine-tuning, why it fits on consumer hardware, and the steps to run your first job.
1 min readRead - Fine-tuning
Ten fine-tuning mistakes I see students make (and made myself)
The mistakes that turn a fine-tuning project into a shrug: no baseline, dirty data, the wrong metric, training on eval data, chasing loss, and skipping the deployment question.
1 min readRead - Fine-tuning
Instruction tuning vs domain adaptation: two different fine-tuning goals
Teaching a model to follow your instructions is not the same as teaching it your domain's language. How the data, the training objective and the evaluation differ, and how to avoid mixing them badly.
1 min readRead - Fine-tuning
From adapter to deployment: merging LoRA weights and serving with vLLM or a Space
What to do after training: keep or merge the adapter, choose between local inference, a vLLM server and a hosted Space, and check that the served model matches what you evaluated.
2 min readRead - Fine-tuning
Small language models (1B–3B): what they're good for after fine-tuning
Where small open models genuinely beat frontier APIs after a fine-tune — classification, extraction, routing, on-device tasks — and where they still fall short, from someone who ships them.
1 min readRead - Fine-tuning
Validate your dataset before you burn a GPU hour: a checklist
A pre-training checklist for fine-tuning datasets: schema, roles, duplicates, leakage, length distribution, label balance and the rendered-template spot check that catches what scripts miss.
1 min readRead - Fine-tuning
Fine-tuning on a laptop or a free GPU: a VRAM budget walkthrough
Where the memory goes when you fine-tune — weights, adapters, optimiser state, activations — and the concrete settings that fit a 1–3B model into 4–6 GB of VRAM without giving up quality.
2 min readRead - Fine-tuning
How to read a loss curve during fine-tuning (and when to stop)
What training and evaluation loss actually tell you, the curve shapes that mean overfitting, a learning rate problem or a broken dataset, and why live telemetry beats reading logs afterwards.
2 min readRead - Fine-tuning
QLoRA hyperparameters that actually matter (rank, alpha, learning rate, epochs)
Which QLoRA settings move results and which are folklore: LoRA rank and alpha, target modules, learning rate, epochs versus dataset size, batch and gradient accumulation, and sane defaults for small models.
2 min readRead - Fine-tuning
Fine-tuning datasets: Alpaca, ShareGPT and chat templates without the confusion
The dataset formats you will meet when fine-tuning an open LLM, how chat templates turn them into tokens, and the formatting mistakes that quietly ruin a training run.
2 min readRead - Fine-tuning
LoRA vs QLoRA vs full fine-tuning: cost, quality and when each makes sense
What LoRA adapters change, what quantisation adds in QLoRA, what full fine-tuning still buys, and a decision guide for students and small teams with limited GPUs.
2 min readRead - Fine-tuning
One base model, five LoRA adapters: serving multiple fine-tuned tasks without five deployments
Swapping LoRA adapters at request time lets one base model serve several fine-tuned behaviours, instead of hosting a separate full model per task.
1 min readRead - Fine-tuning
Fine-tuning checkpoints: what to actually keep, and what to safely delete
Saving every epoch's checkpoint fills a disk fast. A practical policy for what to keep, informed by what a loss curve and eval score actually tell you.
1 min readRead - Fine-tuning
DPO vs supervised fine-tuning: do student projects need preference tuning at all?
Direct Preference Optimization is having a moment. For most student projects, plain supervised fine-tuning on good examples still gets you further, faster.
1 min readRead - Fine-tuning
Fine-tuning vs a longer prompt: a real cost comparison
A well-engineered prompt with examples can match a fine-tune for far less upfront cost — until volume changes the math. Where the crossover actually sits.
1 min readRead - Fine-tuning
Quantizing a fine-tuned model: GGUF, AWQ and GPTQ, and which to pick
A fine-tuned model that doesn't fit your deployment budget needs quantization. The practical differences between the three common formats, in one place.
1 min readRead - Fine-tuning
Fine-tuning with 200 examples: what's actually possible
You don't need ten thousand labelled examples to fine-tune usefully. A realistic plan for what 200-2,000 examples can and can't teach a model.
1 min readRead - Fine-tuning
Picking a base model to fine-tune in 2026: the checklist I actually use
License, size, context length, existing instruction-tuning and community tooling — the factors that matter more than a leaderboard rank when choosing a base model.
1 min readRead - Fine-tuning
Fine-tuning for classification vs generation: two different problems wearing the same name
A sentiment classifier and a chatbot both get called 'fine-tuning', but the data format, evaluation and even the model choice differ sharply between them.
1 min readRead - Fine-tuning
Catastrophic forgetting: why your fine-tuned model got worse at things you didn't train it on
A model fine-tuned hard on one narrow task can lose general capability it had before. What causes it and the simplest ways to prevent it in a student project.
1 min readRead - Fine-tuning
Synthetic data for fine-tuning: when it works and when it teaches the model your own mistakes
Generating training examples with a bigger model can fill a data gap fast, or quietly bake in that model's biases and errors. How to tell which you're getting.
1 min readRead
03Vision & speech31 articles
- Vision & speech
Face recognition that never uploads a face — building FaceVision
FaceVision runs detection, recognition and liveness checks entirely in the browser with ONNX Runtime Web. The backend never sees an image, only vectors.
1 min readRead - Vision & speech
Every page and every word, searchable — building the OCR & Speech Workspace
A Mistral-powered workspace for page-batched PDF OCR, live speech-to-text, and document-scoped chat with citations.
1 min readRead - Vision & speech
Generating alt text with vision AI: a genuinely useful accessibility application
Auto-generating image alt text with a vision-language model is one of the more directly useful, low-risk applications of the technology. What makes a generated alt text good.
1 min readRead - Vision & speech
Video summarization with vision-language models: what's actually feasible today
Full automatic video summarization is an active research area. A realistic, achievable version of it using frame sampling plus a vision-language model.
1 min readRead - Vision & speech
Building a plant or object identifier app: a well-scoped vision project
A photo-to-identification app is visual, has a clear success metric, and is achievable with a pretrained model plus a well-curated reference dataset.
1 min readRead - Vision & speech
Vision API costs: cloud vision APIs vs self-hosting a model, a real comparison
Per-image API pricing looks cheap at low volume and expensive fast. A concrete break-even framework for when self-hosting a vision model pays off.
1 min readRead - Vision & speech
Synthetic image data: when it's a legitimate fix for a scarce dataset
Real labelled images for a narrow vision task are often scarce. Rendered or generated synthetic data can fill the gap — with real limits worth knowing upfront.
1 min readRead - Vision & speech
Vision Transformers, explained for practitioners who just want to use one well
You don't need the full derivation to use a ViT effectively. The practical mental model: patches as tokens, and what that implies for how you feed it images.
1 min readRead - Vision & speech
Barcode and QR detection in noisy real-world images: what actually helps
A QR scanner that works perfectly in a demo often fails on a crumpled receipt or a glare-heavy photo. Practical fixes that improve detection in messy conditions.
1 min readRead - Vision & speech
Document tampering detection: what a vision model can and can't catch
Detecting an altered ID card or edited document is a real, hard vision problem. What signals are checkable today, and what's still beyond a straightforward pipeline.
1 min readRead - Vision & speech
Testing a vision model for bias across skin tones before you ship it
Face and vision models trained on unbalanced datasets perform unevenly across skin tones. What to test for, deliberately, before trusting a model in production.
1 min readRead - Vision & speech
Building a simple attendance system with face recognition: a project walkthrough
A classic, genuinely useful student project. What the pipeline actually needs — enrolment, matching, liveness — and where most attempts cut corners that matter.
1 min readRead - Vision & speech
Privacy by design for AI apps: what I do differently after FaceVision
Practical privacy engineering for AI products: data minimisation, on-device inference, embeddings over raw data, retention and deletion, consent that means something, and how the DPDP Act shapes the defaults.
2 min readRead - Vision & speech
Multimodal LLMs: five student project ideas beyond 'chat with a PDF'
Project ideas that use vision, speech and text together — with the architecture, the hard part and the evaluation for each — chosen to be finishable in a semester and worth showing to a recruiter.
1 min readRead - Vision & speech
Document AI for invoices: schema-first extraction that auditors can trust
Why free-form 'AI reads your invoice' fails, and how schema-validated extraction with deterministic scoring, field-level confidence and comparison against source documents makes document intelligence defensible.
1 min readRead - Vision & speech
Real-time speech-to-text: architecture of a live transcription feature
Streaming audio from the browser over WebSockets, chunking for low latency, partial versus final transcripts, a post-recording accuracy pass, and the UX details that make live captions usable.
1 min readRead - Vision & speech
Designing an OCR pipeline: from scanned PDF to searchable, chunked text
How to turn a 400-page scanned PDF into text you can search and cite: page batching, concurrency, layout awareness, quality checks, and storing page numbers for citations.
1 min readRead - Vision & speech
On-device and browser AI: why it matters for privacy, not just speed
Running inference in the browser with ONNX Runtime Web isn't just a latency trick — it removes an entire category of data breach.
1 min readRead - Vision & speech
Liveness detection basics: stopping a photo from unlocking a face check
Passive and active liveness, what attacks each stops, how to combine a small anti-spoofing model with motion challenges in the browser, and how to measure false accepts honestly.
1 min readRead - Vision & speech
How face recognition works: embeddings, thresholds and why you never store faces
Detection, alignment, embedding and matching explained without the maths, plus how to choose a similarity threshold, handle false matches, and design storage so a breach leaks numbers, not faces.
2 min readRead - Vision & speech
Running ML models in the browser with ONNX Runtime Web: a practical guide
How to ship a vision model that runs entirely in the browser: exporting to ONNX, choosing WebGPU or WASM, pre- and post-processing pitfalls, and verifying the model against its reference implementation.
2 min readRead - Vision & speech
Deploying a vision model to the edge: quantization and ONNX in practice
Getting a vision model to run on a phone or low-power device without a cloud round trip: what quantization and ONNX actually buy you, concretely.
1 min readRead - Vision & speech
Five camera-quality problems that quietly break a vision model in production
A model trained on clean, well-lit images meets a real phone camera and fails in ways the demo never showed. What to test for before deployment.
1 min readRead - Vision & speech
Evaluating a vision model beyond accuracy: the metrics that actually predict production behaviour
A single accuracy number hides which errors a vision model makes and how confidently it makes them. A short list of what to check before shipping.
1 min readRead - Vision & speech
Real-time object detection on a budget GPU: what's actually achievable
A realistic frames-per-second budget for object detection on the free or low-cost GPUs students actually have access to, and how to hit it.
1 min readRead - Vision & speech
Face-matching thresholds: the trade-off between letting the wrong person in and locking the right one out
Every face-matching system picks a similarity threshold. Set it wrong in either direction and you get a security hole or a support-ticket generator.
1 min readRead - Vision & speech
Building a receipt scanner with a vision LLM: a project walkthrough
Extracting merchant, date, line items and total from a photographed receipt using a multimodal model, plus the validation layer that makes it trustworthy.
1 min readRead - Vision & speech
CLIP vs ViT vs a multimodal LLM: picking a vision model without guessing
Three different families of vision model solve overlapping but distinct problems. A decision guide based on what you're actually trying to build.
1 min readRead - Vision & speech
Video frame sampling: you don't need every frame, and processing every frame will bankrupt your project
Running a vision model on every frame of a video is slow and expensive. A practical sampling strategy that catches what matters without the cost.
1 min readRead - Vision & speech
Handwriting recognition for Indian forms: what makes it harder than printed-text OCR
Mixed scripts, inconsistent handwriting and low-quality scans make handwritten form recognition a genuinely different problem from OCR on printed text.
1 min readRead - Vision & speech
Image embeddings for visual search: how 'find similar' actually works
The mechanics behind a 'find visually similar items' feature: embedding images into a shared vector space and searching it the same way text RAG does.
1 min readRead
04Shipping AI33 articles
- Shipping AI
Handling long-running AI agents: timeouts, checkpoints and partial progress
An agentic task that runs for minutes needs a different failure model than a single API call. Timeouts, checkpointing and resuming partial work, done right.
1 min readRead - Shipping AI
Building an internal AI tool for your own team: a lower-stakes way to ship something real
An internal tool has real users (your teammates) without the stakes of a public product. A genuinely good way to practice the full build-to-deploy loop.
1 min readRead - Shipping AI
Versioning prompts like you version code: a simple system that pays off fast
A prompt edited directly in the codebase with no history becomes impossible to debug when quality regresses. Simple prompt versioning fixes this cheaply.
1 min readRead - Shipping AI
Reading an invoice like an auditor — building DocuLens AI
DocuLens AI extracts, compares and compliance-checks business documents with schema-validated, deterministic scoring — not a free-text LLM summary.
1 min readRead - Shipping AI
Graceful degradation: what your app should do when the LLM provider is down
Every provider has occasional outages. An app with no fallback plan goes fully dark; one designed for degradation keeps working in a reduced but honest way.
1 min readRead - Shipping AI
Setting up cost alerts for AI APIs before the bill surprises you
A rate limit stops a runaway loop from crashing your app. A cost alert stops it from crashing your budget. Both are needed, and most student projects only build the first.
1 min readRead - Shipping AI
Webhooks and async callbacks for AI pipelines: notifying instead of polling
Polling a job-status endpoint every few seconds wastes resources on both sides. A webhook-based callback pattern is often the better fit for slow AI tasks.
1 min readRead - Shipping AI
Building a support-ticket triage system with an LLM: a project walkthrough
Classifying, prioritising and routing incoming support tickets is a genuinely useful, well-scoped AI project with a clear before-and-after metric.
1 min readRead - Shipping AI
Shipping side projects that survive demo day
Why most student projects die the day after they are shown, and the handful of habits that keep one alive: a real user, a deploy that stays up, a changelog, monthly maintenance, and knowing when to archive well.
2 min readRead - Shipping AI
Handling personal data (PII) in LLM applications without it leaking where it shouldn't
Sending user PII to a third-party LLM API, logging it in plaintext, or letting it leak into a training set are three distinct risks. What to actually do about each.
1 min readRead - Shipping AI
A/B testing an AI feature: what to actually measure beyond click-through rate
Standard A/B testing metrics don't capture whether an AI answer was actually correct or useful. What metrics matter specifically for testing AI feature changes.
1 min readRead - Shipping AI
Agentic AI patterns: tools, memory and guardrails without the hype
What an 'agent' is in production terms — a loop with tools — and the patterns that make one safe to ship: least-privilege tools, retrieved text as data, bounded loops, inspectable memory and a pause path.
2 min readRead - Shipping AI
Feature flags for AI features: rolling out a new model or prompt safely
Shipping a new prompt or model version to everyone at once is how a regression reaches every user simultaneously. Feature flags let you roll out gradually and roll back fast.
1 min readRead - Shipping AI
Secrets management for student projects: never commit an API key again
The habits that keep keys out of repositories and demos: ignored environment files, committed example files, server-side keys only, rotation after a leak, and what to do in the first ten minutes if one gets out.
1 min readRead - Shipping AI
Controlling LLM costs: token budgets, caching and cheaper routing
How to keep an AI product's bill predictable: per-request context caps, embedding and response caches, routing simple requests to small models, fallback spend limits, and the dashboard that shows cost per feature.
1 min readRead - Shipping AI
Building SCULT Studio, the internal CRM our own agency runs on
As the GenAI engineer at SCULT INDIA, I designed and built the internal platform for project, client and operations management that the whole team uses daily.
1 min readRead - Shipping AI
Testing LLM applications: unit tests, golden sets and regression prompts
How to test software whose core component is non-deterministic: test the deterministic parts hard, hold golden prompts with expected properties, run an evaluation set as a test, and gate deploys on thresholds.
2 min readRead - Shipping AI
FastAPI + Next.js: the architecture I teach for every AI product
The two-service architecture behind the AI apps Pranjul Rathour ships and teaches in workshops: a Python API for models and retrieval, a Next.js front end for users, and the seams that keep both simple.
1 min readRead - Shipping AI
From notebook to production: a checklist for shipping AI apps
Docker, CI, structured logging, tests — the unglamorous parts that separate a demo from something that stays up.
1 min readRead - Shipping AI
Choosing an LLM provider: a decision framework
Groq, Claude, Gemini, OpenAI — the right answer depends on your constraints, not on whichever model topped a benchmark this month.
1 min readRead - Shipping AI
Prompt injection vs memory poisoning: a practical guide for builders
Prompt injection resets when a chat ends. Memory poisoning doesn't. If your agent has memory, it needs a different threat model.
1 min readRead - Shipping AI
Building a RAG app on a free-tier server: the memory-budget playbook
How to fit a production AI service inside a 512 MB free instance — the choices that matter and the ones that don't.
1 min readRead - Shipping AI
Deterministic document extraction: why "AI reads your invoice" needs guardrails
An LLM asked the same compliance question twice can give two different answers. Here's how to build a document-AI system that doesn't.
1 min readRead - Shipping AI
Monitoring an AI app in production on a student budget
You don't need an enterprise observability stack to know when your AI feature is failing. A minimal, free monitoring setup that catches the failures that matter.
1 min readRead - Shipping AI
CI for AI projects: what's actually worth testing automatically
You can't unit-test 'is this a good answer', but you can test the parts around it. What continuous integration should cover in an LLM-backed project.
1 min readRead - Shipping AI
Writing a README that actually gets your AI project looked at
A recruiter or judge spends thirty seconds on your README before deciding whether to look further. What that thirty seconds needs to show.
1 min readRead - Shipping AI
Environment variables and config for AI apps: getting API keys right from day one
Hardcoded API keys are the single most common security mistake in student AI projects. A simple config setup that avoids it permanently.
1 min readRead - Shipping AI
Deploying a Python AI app for free: a realistic path from laptop to a working URL
You don't need a paid server to put a working AI demo online. A practical path using free tiers, with the trade-offs each one carries.
1 min readRead - Shipping AI
Error handling for flaky LLM APIs: what to do when the provider times out mid-demo
LLM provider APIs occasionally time out, rate-limit, or return malformed output. A practical error-handling pattern that keeps your app usable anyway.
1 min readRead - Shipping AI
Caching LLM responses safely: the checklist before you cache anything
Caching cuts cost and latency, but caching the wrong thing serves stale or wrong answers confidently. What's safe to cache and what isn't.
1 min readRead - Shipping AI
Background jobs for slow AI tasks: don't make the user wait on a 30-second request
Some AI tasks — a long document summary, a batch classification — take too long for a synchronous HTTP request. A simple background-job pattern that fixes it.
1 min readRead - Shipping AI
Logging and tracing LLM calls: what a debuggable AI app actually needs
When an AI feature misbehaves in production, print statements won't save you. A practical logging and tracing setup for LLM-backed applications.
1 min readRead - Shipping AI
Rate-limiting your own AI API before your users (or a bug) do it for you
One runaway frontend loop can burn a month's API budget in an hour. A simple, practical rate-limiting setup for a student AI project's backend.
1 min readRead
05Hackathons47 articles
- Hackathons
Running a judges-only debrief right after the event: the fifteen minutes organisers skip
A short debrief among judges, right after results are decided, surfaces process improvements while they're still fresh — and is skipped far more often than it should be.
1 min readRead - Hackathons
What I wish someone had told me before my first hackathon judging gig
First-time judging comes with a few surprises nobody warns you about — the pace, the fatigue, the awkward moments. A few things worth knowing in advance.
1 min readRead - Hackathons
Judging a hackathon where you're also personally sponsoring a prize
Occasionally a judge is also the one putting up a prize from their own pocket or company. How to keep that arrangement transparent and fair to every team.
1 min readRead - Hackathons
How organisers should train first-time judges before the event, not during it
A briefing on event day is necessary but not sufficient. What organisers can do beforehand to prepare judges who've never done this before.
1 min readRead - Hackathons
Hackathon sponsors and judging: staying neutral when a sponsor's tech is on the table
Sponsor prizes for 'best use of X API' create a real conflict-of-interest question for judges. How to keep sponsor-track judging fair and clearly separated.
1 min readRead - Hackathons
Annapurna: from a 36-hour hackathon prototype to a real food-sharing platform
How a hackathon-winning project — a real-time platform connecting food donors, NGOs and volunteers — became something we kept building after the trophy.
1 min readRead - Hackathons
Beating 15+ teams at BYTEBATTLE with Annapurna
Team TechVerse's 1st prize at MeetKats' BYTEBATTLE hackathon, and what it took to win against 60+ competing developers.
1 min readRead - Hackathons
Three times I didn't win first place (and what stuck more than the wins)
A top-3 finish at IIT Kanpur, a cybersecurity runner-up and a debate runner-up — the near-misses that taught me more than the trophies.
1 min readRead - Hackathons
Pitching BrandHive to a Vice Chancellor — winning Product Genesis at CSJMU
1st prize at Vividhotsava 2025's Product Genesis for BrandHive, a startup concept helping small businesses with AI content and UGC collaborations.
1 min readRead - Hackathons
Winning Changethon 2025 at IIT Roorkee with an AI agribot
How Team Techverse won 1st prize at IIT Roorkee's National Social Summit with Krishi Gyan, an AI agribot for real-time farming guidance.
1 min readRead - Hackathons
Building a personal judging checklist you'll actually use pitch after pitch
A rubric sets the criteria; a personal checklist is what keeps a judge consistent across the tenth pitch of a long day, not just the first.
1 min readRead - Hackathons
What I look for in a two-minute lightning pitch (it's not what a ten-minute pitch rewards)
A lightning round compresses everything. The criteria that matter in a ten-minute pitch don't transfer directly — here's what actually separates entries at two minutes.
1 min readRead - Hackathons
Judging alongside non-technical co-judges: making a mixed panel actually work
A panel with a technical judge and a business or design judge scores differently by default. How to align without flattening what each judge brings.
1 min readRead - Hackathons
When a team pushes back on your score: how to handle it without damaging trust
Occasionally a team disagrees, sometimes loudly, with a judge's score. How to respond in a way that protects both the team's dignity and the process's integrity.
1 min readRead - Hackathons
Etiquette at the judges' table during live demos: what helps and what undermines a team
Small things a judge does during a live demo — phone use, side conversations, facial reactions — visibly affect a nervous team's performance. What to actually do instead.
1 min readRead - Hackathons
Hackathon judging vs startup pitch judging: two different jobs that look similar
Both involve a panel and a pitch, but a hackathon rewards a working 36-hour build; a startup pitch rewards a viable business. Judging one like the other misfires.
1 min readRead - Hackathons
Judging a solo hacker's submission fairly against team submissions
A solo participant is competing against teams of three or four with more hands and hours. Should the rubric adjust for that, and if so, how?
1 min readRead - Hackathons
Becoming a repeat hackathon judge: what organisers actually remember about you
Organisers rarely remember a judge's exact scores. What they do remember — punctuality, written feedback quality, calm under a tight schedule — decides the next invite.
1 min readRead - Hackathons
What a hackathon judge should include in a post-event report to organisers
Beyond the scores, a short report back to organisers on rubric gaps, standout teams and process friction helps the next event run better than this one.
1 min readRead - Hackathons
Judging a virtual hackathon with teams across time zones
A global online hackathon spreads teams across a dozen time zones. What that actually changes about scheduling, fatigue and fairness for judges.
1 min readRead - Hackathons
Losing a hackathon well: what to keep, what to ditch
How to lose a hackathon and come out ahead: reading the judges' feedback, separating the pitch problem from the product problem, and why STROT taught more than three first prizes.
2 min readRead - Hackathons
STROT: the hackathon I didn't win, and the pitch lesson that mattered more
We lost this one with a two-minute pitch for a system that took days to build. Clarity beat depth, and that lesson has outlasted every trophy.
1 min readRead - Hackathons
Judging AI hackathon projects: what to check when every team says 'we used AI'
A judge's checklist for AI projects: distinguishing a real inference from a hard-coded response, asking about evaluation and failure, checking data handling, and scoring the human decisions around the model.
2 min readRead - Hackathons
The questions judges ask in hackathon Q&A, and how to answer them
Fifteen questions that come up at nearly every student hackathon judging table — on the demo, the technology, the user and the team — with what a strong answer sounds like and what a weak one reveals.
2 min readRead - Hackathons
What a hackathon mentor actually does (and how to use one)
The mentor's job during a hackathon: unblock without building, ask the scoping question, run a mock pitch, and know when to step back — plus how teams should use the mentor hours they usually waste.
1 min readRead - Hackathons
How hackathon judges should give feedback (so students actually use it)
Feedback that changes what a team does next: tie it to a criterion, name one strength and one fix, separate pitch from product, and write it down — the practice Pranjul Rathour brings to student panels.
2 min readRead - Hackathons
How to run a college hackathon students remember: an organiser's guide
A practical guide for student organisers and faculty coordinators: problem statements with users, judges and mentors worth inviting, a schedule with sleep in it, and the feedback loop that makes teams come back.
2 min readRead - Hackathons
Your first hackathon as a first-year: how to be useful with limited skills
A first-year's guide to being valuable on a hackathon team before you can code well: user research, the pitch, testing, seeded data, and the habits that get you invited back.
1 min readRead - Hackathons
Designing hackathon problem statements that judges can actually score fairly
A vague problem statement produces incomparable projects. Writing problem statements with the rubric already in mind fixes judging before it starts.
1 min readRead - Hackathons
What a judges' briefing should actually cover (fifteen minutes, before the first pitch)
A short calibration meeting before judging starts prevents most scoring inconsistency. What that fifteen minutes should actually contain.
1 min readRead - Hackathons
After the hackathon: turning a prototype into a product people use
What to do in the two weeks after a hackathon — win or lose — to turn a 36-hour prototype into a project with users: the rewrite decision, the first ten users, and what to keep from the demo.
1 min readRead - Hackathons
Conflicts of interest in hackathon judging: what to do when you know a team
A judge who mentored a team, taught them, or knows them personally has a real conflict. How I've seen it handled well, and what organisers should ask upfront.
1 min readRead - Hackathons
Hackathons in India for students: where to look and how to pick
Where student hackathons are announced, how to tell a serious event from a certificate mill, what to check about judging and prizes, and how to plan a season of two or three good events.
1 min readRead - Hackathons
Tie-breaking in hackathon judging: a process that doesn't feel arbitrary to the teams
A coin flip or an organiser's gut call at the top of the leaderboard erodes trust fast. A tie-breaking process decided before the event, not during it.
1 min readRead - Hackathons
How many teams can one judge fairly see in a day? The number organisers underestimate
Judging quality drops sharply after a certain number of pitches in a day. A realistic per-judge cap and how to schedule around it.
1 min readRead - Hackathons
25 AI hackathon project ideas for 2026 (with the hard part of each)
Project ideas across agriculture, education, health, accessibility, campus life and small business — each with the one hard problem that makes it worth a judge's attention.
2 min readRead - Hackathons
Judging a remote hackathon vs an on-site one: what actually changes
A demo over a video call fails differently than one in a room. What a judge needs to adjust for a remote hackathon, and what stays exactly the same.
1 min readRead - Hackathons
Five mistakes every first-time hackathon judge makes (I made most of them too)
Scoring on charisma, comparing teams instead of scoring against the rubric, and other first-time-judge mistakes — and the simple fixes for each.
1 min readRead - Hackathons
A hackathon pitch deck template that fits in three minutes
Seven slides, thirty seconds each, in the order judges need to hear them — with what to say on each, what to cut, and how to hand off to the live demo without losing the room.
1 min readRead - Hackathons
Innovation vs execution: how I actually weigh a hackathon rubric's criteria
Most rubrics list innovation, execution, impact and presentation with no guidance on relative weight. A concrete weighting scheme and the reasoning behind it.
1 min readRead - Hackathons
Panel judging vs solo judging: why a hackathon needs more than one judge in the room
A single judge's taste becomes the whole competition's outcome. A panel with a shared rubric and a calibration round fixes that — here's how the two formats actually differ.
1 min readRead - Hackathons
A 36-hour hackathon timeline, hour by hour
A realistic schedule from problem selection to judging — with the two checkpoints that separate finished projects from ambitious fragments, and where to sleep.
1 min readRead - Hackathons
How to become a hackathon judge: what actually gets you invited to the panel
Organisers don't search a directory for judges — they ask people they already trust. What builds that trust before the first invitation ever arrives.
1 min readRead - Hackathons
Forming a hackathon team: roles, skills and the mistake most teams make
How to assemble a team that finishes: the four roles that matter, why five friends with the same skill lose, how to split work in the first hour, and what mentors notice about team dynamics.
1 min readRead - Hackathons
The hackathon demo that works live: a technical checklist
How to make a demo survive the judging table: seeded data, a rehearsed path, offline fallbacks, no live deploys in the last hour, and what to say when something breaks anyway.
2 min readRead - Hackathons
A hackathon judging rubric organisers can copy
A complete, weighted rubric for student hackathons with scoring descriptors for each level, notes on calibrating judges, and how to publish criteria so teams build toward them.
2 min readRead - Hackathons
What hackathon judges actually look for: a rubric from both sides of the table
The four things a judge is scoring in a three-minute pitch, why working demos beat ambitious slides, and the rubric Pranjul Rathour brings to a student hackathon panel.
2 min readRead
06Mentoring & speaking57 articles
- Mentoring & speaking
Speaking about AI to non-engineering college audiences: what actually needs to change
A talk built for CS students falls flat with commerce or arts students. What genuinely needs adjusting, and what stays honest and useful either way.
1 min readRead - Mentoring & speaking
Building a personal brand as a campus speaker, without it feeling like self-promotion
Consistent positioning — a clear niche, a recognisable set of talks — helps organisers know exactly what to book you for. How to do this without it feeling hollow.
1 min readRead - Mentoring & speaking
Mentoring through open source: a different channel from talks and workshops
Reviewing pull requests and answering issues on a public repo is its own form of mentorship — quieter, asynchronous, and reaching people a talk never would.
1 min readRead - Mentoring & speaking
Speaking at international student conferences remotely: what changes across borders
A remote talk to students in another country adds time zones, cultural context and sometimes language considerations on top of the usual remote-speaking mechanics.
1 min readRead - Mentoring & speaking
My mentoring philosophy: create more than you consume
The one line I repeat in almost every session, and why it's the most actionable advice I have for a student who feels behind.
1 min readRead - Mentoring & speaking
Judge, mentor, or speaker — which one does your event actually need?
The three roles look similar from the outside but need completely different preparation. Here's how to tell which one your event needs.
1 min readRead - Mentoring & speaking
How I structure a free campus workshop (and why colleges should book one)
What a hands-on session with me actually looks like on the ground: live system, real repo, one takeaway students can act on that week.
1 min readRead - Mentoring & speaking
Building credibility as a speaker before you've ever given a college talk
Nobody gets a first invite from nowhere. What to build first — publicly — so a fest team has a reason to trust an unproven speaker with their audience.
1 min readRead - Mentoring & speaking
Sharing a stage: co-speaking with another guest at the same college event
A fest sometimes books two speakers for the same slot or back to back. Coordinating with the other speaker beforehand turns overlap into a stronger combined session.
1 min readRead - Mentoring & speaking
When a college cancels your session at the last minute: how to handle it well
Fest schedules slip and sessions get cancelled or moved with little notice. A gracious response protects the relationship for next time, far more than the cancellation itself matters.
1 min readRead - Mentoring & speaking
Why I judge a hackathon pitch on storytelling as much as code
Three hackathon wins and a loss on a two-minute pitch taught me the same thing from opposite sides of the judging table.
1 min readRead - Mentoring & speaking
Turning a single guest lecture into an ongoing mentor relationship with a student
Most guest lectures end when the session does. A few students always ask a follow-up question — what happens after that first message decides everything.
1 min readRead - Mentoring & speaking
Mentoring through a college's alumni network instead of one-off visits
Some colleges have a standing alumni-mentor structure. Plugging into it, rather than only doing one-off guest talks, changes the kind of impact you can have.
1 min readRead - Mentoring & speaking
Why I speak at colleges for free
The economics of free campus talks, and why I still say yes to almost every student community that asks.
1 min readRead - Mentoring & speaking
How I keep track of every college I've spoken at (and why it matters more than it sounds)
A simple spreadsheet of college, date, topic and outcome prevents repeated talks, missed follow-ups, and forgotten relationships as the number of sessions grows.
1 min readRead - Mentoring & speaking
The speaker kit: photos, video clips and copy organisers actually ask for
A promo poster needs a decent photo and a short bio within the hour it's requested. Having a ready speaker kit turns a same-day ask into a five-minute reply.
1 min readRead - Mentoring & speaking
Building an ongoing relationship with several college tech clubs at once
One-off talks are fine. A standing relationship with a handful of clubs — where they come back every semester — compounds far more for both sides.
1 min readRead - Mentoring & speaking
Saying yes to every college invite (and when I've learned to say no)
Early on, every invite feels precious. A few criteria for declining without guilt, once speaking becomes a repeat commitment rather than a rare opportunity.
1 min readRead - Mentoring & speaking
Guest of honour at a college fest: what the role actually involves beyond a ribbon-cutting
A guest-of-honour slot is often assumed to be ceremonial. Done well, it's a short talk, real interaction with students, and a few photos that matter to the organisers.
1 min readRead - Mentoring & speaking
From giving one nervous talk to becoming someone colleges invite back: what actually changed
The gap between a first talk and a sustained speaking practice isn't talent — it's a handful of specific habits built up one invitation at a time.
1 min readRead - Mentoring & speaking
Speaking at several college fests in one month without burning out
Fest season clusters invitations into a few busy weeks. A realistic system for saying yes to more without every talk becoming a rushed, tired version of itself.
1 min readRead - Mentoring & speaking
What makes a college invite the same speaker back the next year
A one-off good talk is nice. A repeat invitation means something specific happened after the first one. What that specific thing usually is.
1 min readRead - Mentoring & speaking
Measuring the impact of a campus session: beyond attendance and feedback forms
How coordinators and speakers can tell whether a talk or workshop worked: the two-week follow-up, shipped-project counts, club sign-ups, and the questions students ask afterwards.
1 min readRead - Mentoring & speaking
Speaking fee vs speaking free: how I actually decide, and why I default to free for students
Not every speaking engagement is priced the same. A transparent, simple framework for when a speaker charges and when they don't — and why.
1 min readRead - Mentoring & speaking
Faculty and practitioners co-teaching GenAI: a model that works for colleges
How departments can pair a faculty member with a visiting practitioner for a GenAI module: who owns theory, who owns the build, how to schedule visits, and what students gain that neither could give alone.
1 min readRead - Mentoring & speaking
Writing a talk abstract that gets accepted, not just submitted
A vague abstract gets lost in a pile of forty submissions. A specific, outcome-led abstract gets picked. The difference is one paragraph, written well.
1 min readRead - Mentoring & speaking
What college fest organisers are actually screening for when they pick a speaker
Reliability and fit for the audience matter more to organisers than raw credentials. What actually moves an organiser from considering to booking.
1 min readRead - Mentoring & speaking
An AI awareness session for non-CS students: what to teach in one hour
A session plan for commerce, arts, science and management students: what AI tools can and cannot do, how to use them honestly for study and work, privacy basics, and one hands-on exercise that lands.
1 min readRead - Mentoring & speaking
Speaking at a virtual conference vs in person: what actually needs to change
The talk content doesn't need to change much between formats — the delivery mechanics absolutely do. What to adjust for a virtual college fest session.
1 min readRead - Mentoring & speaking
A year of student community events: a calendar template for tech clubs
A twelve-month plan for a college tech club or community: sessions, hackathons, speaker visits, project demo days and the quiet months — with who owns what, adapted from TechVerse Enclave.
1 min readRead - Mentoring & speaking
Handling Q&A after a technical talk: the questions that actually matter more than the talk
A talk is scripted; Q&A isn't. How to handle the three kinds of questions a technical audience actually asks, without losing the room.
1 min readRead - Mentoring & speaking
Planning an induction seminar for first-years: what to cover and what to skip
For departments and coordinators: how to design an induction talk that first-years remember — the three-year roadmap, one honest failure story, a concrete first task, and the mistakes that lose the room.
1 min readRead - Mentoring & speaking
The speaker one-pager: what a fest team actually wants before they book you
A single page — topics, formats, past sessions, logistics, contact — cuts a three-email back-and-forth down to one reply. What to put on it.
1 min readRead - Mentoring & speaking
A 90-minute GenAI guest lecture outline that faculty can slot into a semester
A ready-to-use guest lecture plan: what a language model is, retrieval and hallucination live, where fine-tuning fits, how production systems stay honest, and a take-home exercise aligned to a course.
1 min readRead - Mentoring & speaking
How to write a speaker bio that actually gets you booked for a college fest
Most speaker bios are a resume in paragraph form. A bio that gets booked answers one question for the organiser: why this speaker, for this audience.
1 min readRead - Mentoring & speaking
What a good mentee does: the habits mentors notice
From the mentor's side: the six habits that make students worth investing in — showing work, closing loops, asking narrow questions, shipping between sessions — and why they matter more than talent.
1 min readRead - Mentoring & speaking
Fests at IITs, NITs and state universities: how speaker outreach actually works
For fest teams and for practitioners: how technical fests find speakers and judges, what a good outreach message contains, timelines, and why smaller colleges often get better sessions.
2 min readRead - Mentoring & speaking
How to ask a mentor for help (so you actually get it)
The message structure that gets replies from busy people: context, what you tried, the specific question, and the one-line ask — with before-and-after examples from real student messages.
1 min readRead - Mentoring & speaking
Why colleges should book practitioners, not only professors, for AI sessions
The case for bringing working engineers onto campus: currency of the toolchain, the language of production, honest career information, and a network students cannot get from a syllabus.
1 min readRead - Mentoring & speaking
Office hours for students: how I run them and why they beat DMs
A weekly open hour for student questions: the format, the rules that keep it useful, what students bring that gets the best answers, and how a community turns one hour into shared knowledge.
1 min readRead - Mentoring & speaking
The ten questions students ask me most (and my honest answers)
The questions that come up after every campus session — on DSA versus projects, which language, hackathons, certificates, startups versus jobs, and whether it is too late — with the answers a mentor actually gives.
2 min readRead - Mentoring & speaking
Mentoring a hackathon team remotely: what works over a call
How to mentor student teams you cannot sit beside: three fixed check-ins, screen-share rules, a shared scoping doc, and how to run a mock pitch over video.
1 min readRead - Mentoring & speaking
How I prepare a 45-minute campus talk
The preparation behind a talk students remember: one question the talk answers, a story with a failure in it, a live demo, three rehearsals with a timer, and the one thing I ask coordinators before I write a slide.
2 min readRead - Mentoring & speaking
Mentorship session, workshop or talk: pick the right format for your students
A guide for coordinators choosing between a guest talk, a hands-on workshop and a mentorship session — what each achieves, how long it needs, who should attend, and how to combine them.
1 min readRead - Mentoring & speaking
Running a remote guest lecture that doesn't feel remote
How to deliver a guest lecture over video to a college classroom: the room setup on their side, the camera and audio on yours, interaction every ten minutes, and a co-host who owns the chat.
2 min readRead - Mentoring & speaking
From campus to community: the TechVerse Enclave playbook
How a college group became a 500+ member developer community: the first thirty members, the weekly rhythm, admins who own areas, sessions that convert lurkers, and what nearly killed it.
2 min readRead - Mentoring & speaking
Talks I offer colleges: RAG, fine-tuning, shipping AI, hackathons and careers
Pranjul Rathour's current talk and workshop menu for colleges and student communities, with the audience, duration and takeaway for each — from a 40-minute induction talk to a full-day GenAI build.
2 min readRead - Mentoring & speaking
A tech club session format that keeps 60 students engaged for 90 minutes
The session structure Pranjul Rathour uses for campus workshops: a live build in front of the room, a fifteen-minute hands-on block, a show-and-tell, and the three rules that stop a workshop becoming a lecture.
1 min readRead - Mentoring & speaking
For placement cells: how industry GenAI sessions improve placements
A guide for training and placement officers: which sessions change interview outcomes, how to schedule them across the year, what to ask speakers for, and how to turn one visit into internships.
2 min readRead - Mentoring & speaking
Inviting a hackathon judge: what organisers should send in the first email
How to invite a judge who will actually improve your hackathon: the brief, the rubric, the schedule, the number of teams, mentoring versus judging, and the small courtesies that make judges say yes again.
2 min readRead - Mentoring & speaking
A one-day GenAI workshop syllabus for colleges (hands-on, no fluff)
A six-hour workshop plan that takes students from zero to a deployed retrieval-augmented app: prerequisites, four build blocks, the evaluation exercise, and what to send students afterwards.
2 min readRead - Mentoring & speaking
Building TechVerse Enclave: lessons from a 500-member developer community
What actually works, and what flops, when you found and grow a student tech community while mentoring 200+ people through it.
1 min readRead - Mentoring & speaking
How to invite a GenAI guest speaker to your college: a checklist for coordinators
Everything a student coordinator or faculty member needs to book a practitioner for a talk or workshop: what to put in the first email, dates, formats, logistics, honorarium norms and follow-up.
1 min readRead - Mentoring & speaking
Mentoring first-year students: what works and what wastes their time
Lessons from mentoring 200+ students through TechVerse Enclave: the three things first-years actually need, the advice that backfires, and a simple structure for a semester of mentoring.
2 min readRead - Mentoring & speaking
What I told 100+ first-years at the VSICS Induction
The three-year roadmap I gave incoming engineering students: foundations and AI tools, then hackathons and internships, then placements or a startup.
1 min readRead - Mentoring & speaking
Teaching AI-powered software development to first-year engineers
Inside the VSICS session where I mapped how AI now touches every stage of the software lifecycle, from requirements to maintenance.
1 min readRead - Mentoring & speaking
What a college debate stage taught me about arguing a position I don't hold
At INSIGHT — The Talk Show, VSICS Kanpur, I argued against the motion on nationalism and policy. It's a soft-skill exercise every engineer should try.
1 min readRead
07Career guides36 articles
- Career guides
Cold-emailing a mentor or speaker (and actually getting a yes)
What makes an invitation email easy to say yes to, from someone who receives them and someone who's sent plenty.
1 min readRead - Career guides
Writing your first technical blog post: a template
A simple structure for turning a project you built into a post worth reading, even if you've never written one before.
1 min readRead - Career guides
How to start a tech community on your campus
TechVerse Enclave went from five admins to 500+ developers. Here's the actual sequence, not the highlight reel version.
1 min readRead - Career guides
NIMCET, placements, or a startup: choosing your Year 3
The three real paths a final-year engineering student has, and an honest framework for choosing between them instead of drifting into one.
1 min readRead - Career guides
Building your personal brand as a CS student
LinkedIn, YouTube and GitHub are your resume now, whether you've decided to use them that way or not. Here's how to do it without it feeling fake.
1 min readRead - Career guides
How to pitch a startup idea in 2 minutes (and why most pitches fail)
I lost a hackathon on a two-minute pitch for something that took two days to build. Here's the structure I use now.
1 min readRead - Career guides
Portfolio over certificates: what recruiters actually look at
A recruiter opens GitHub before they open your certificate folder. Here's how to build a portfolio that survives that first thirty seconds.
1 min readRead - Career guides
How to win your first hackathon: a practical playbook
Three first-place wins, a top-3 finish and one loss on a two-minute pitch — the real pattern behind winning, condensed into a playbook.
1 min readRead - Career guides
AI tools every CS student should learn in 2026
Not a hype list. The tools that actually change what a student can ship in a semester, and how to use them without becoming dependent on them.
1 min readRead - Career guides
Building a tech career from Kanpur: communities, events and what I'd do again
A local guide for students in Kanpur and similar cities: the communities and events that matter, how to get into hackathons at IITs and universities nearby, and how it led to building production AI.
2 min readRead - Career guides
Choosing a final-year AI project that a recruiter will care about
How to pick and scope a final-year project in AI that is finishable, defensible in a viva and useful in an interview: real users, an evaluation plan, one hard part, and the documentation that makes it count.
2 min readRead - Career guides
Coding contests vs hackathons: which one helps your career more?
An honest comparison for students choosing where to spend weekends: what contests like TCS CodeVita prove, what hackathons prove, how recruiters read each, and a plan that uses both.
1 min readRead - Career guides
Cold outreach to startups for internships: messages that got replies
How to write to a founder or engineer you have never met and get a reply: research that shows, one link, one specific offer, brevity, and the follow-up — with templates students can adapt.
1 min readRead - Career guides
Building in public as a student: a 30-day plan
A month-long plan for shipping one project publicly: daily commits, three progress posts, a demo video, a write-up of what broke, and how building in public turns into mentors, teammates and invitations.
1 min readRead - Career guides
Getting a remote GenAI role from a tier-2 city in India
How students and freshers outside the metro hiring loop land remote AI engineering work: where roles are posted, what remote employers screen for, asynchronous communication as a skill, and the portfolio that travels.
1 min readRead - Career guides
Open source for students: contributing to AI repos without feeling lost
How to make a first meaningful contribution to an AI or developer-tools repository: choosing a project you use, reading before writing, documentation and test contributions, and turning issues into a track record.
1 min readRead - Career guides
Learning Python for AI in 2026: the parts that matter
The Python a GenAI engineer actually uses — data structures, async, typing, packaging, reading library code — and the parts students over-invest in, with a four-week plan built around a real project.
1 min readRead - Career guides
Internship, freelance or startup in Year 2?
How a second-year student should choose between an internship, freelance client work and starting something — what each teaches, what each costs, and the combination that worked for the team behind SCULT INDIA.
2 min readRead - Career guides
LinkedIn for CS students: posting projects without cringe
How to use LinkedIn as a student engineer: what to post, how often, the post structure that gets replies from practitioners, and what to avoid — from someone whose posts led to talks, judging invitations and clients.
1 min readRead - Career guides
A GitHub profile recruiters actually read
How to arrange a GitHub profile so a reviewer finds the right three projects in thirty seconds: pins, a profile README with numbers, project READMEs that lead with the problem, and what to archive.
1 min readRead - Career guides
A resume for AI roles as a fresher: what to put above the fold
How to structure a one-page resume for GenAI and AI engineering roles when you have no full-time experience: projects with numbers first, a skills line that matches job posts, and what to cut.
1 min readRead - Career guides
From a non-IIT college to AI engineer: a realistic path
What an engineering student in a tier-2 city can do to reach a GenAI engineering role: the syllabus gaps and how to fill them, where the opportunities are, and Pranjul Rathour's own path from VSICS Kanpur.
2 min readRead - Career guides
Five AI portfolio projects that get interviews (and five that don't)
From reviewing hundreds of student portfolios: the projects that make a reviewer lean forward, the ones that get skipped, and the three things that turn any project into an interview.
1 min readRead - Career guides
GenAI engineer interview questions I'd ask (with what a good answer sounds like)
Twenty questions across RAG, fine-tuning, evaluation, reliability and product judgement that separate candidates who built systems from candidates who read about them — with the shape of a strong answer for each.
2 min readRead - Career guides
GenAI engineer roadmap 2026 for Indian students
A twelve-month, project-first roadmap to a GenAI engineering role from any college: Python and APIs, a RAG project, a fine-tuning project, a shipped product, and the evidence recruiters actually check.
2 min readRead - Career guides
From a college project to a startup: what actually changes, beyond the pitch deck
Turning a hackathon project into a real startup changes the constraints far more than it changes the code. What founders wish they'd known first.
1 min readRead - Career guides
Time management for CS students trying to ship projects and pass exams at the same time
Side projects and coursework compete for the same hours. A realistic weekly structure that protects both instead of sacrificing one for the other.
1 min readRead - Career guides
Side project ideas that actually teach production skills, not just model-calling
Calling an LLM API is the easy 10% of a real AI product. A list of side-project ideas chosen specifically to force the harder 90%.
1 min readRead - Career guides
TCS CodeVita 2025: complete preparation and registration guide
How to register for TCS CodeVita, what the coding rounds actually test, and how to prepare no matter which degree you're in.
1 min readRead - Career guides
How to read an AI research paper as a student without drowning in the math
A practical reading order for a dense paper: abstract, figures, conclusion, then the method — and when it's fine to skip the proofs entirely.
1 min readRead - Career guides
Building a 'second brain' for learning AI: what to actually write down
The GenAI field moves fast enough that notes decay within months. A practical note-taking system that stays useful instead of becoming a graveyard.
1 min readRead - Career guides
Remote work etiquette for a student's first remote internship or job
Remote work has unwritten rules nobody explains to a first-time remote employee. The habits that make a manager trust you were actually working.
1 min readRead - Career guides
Negotiating your first tech offer as a student: what's actually negotiable
Most first-time offers have more room than students assume, but not in the places students usually push on. A realistic guide to what to ask for.
1 min readRead - Career guides
Answering 'tell me about a project' without rambling for five minutes
The most common interview question also gets the worst answers — a chronological walkthrough with no point. A structure that fixes it in one pass.
1 min readRead - Career guides
What recruiters actually search for on LinkedIn (and why your headline matters more than your summary)
Recruiters search by keyword before they ever read a profile. What that means for how a student engineer should write their headline and skills section.
1 min readRead - Career guides
Technical interview prep for GenAI roles: what's actually being tested
GenAI interviews mix classic DSA with system-design questions about retrieval, evaluation and cost. A realistic prep plan for both halves.
1 min readRead
08Journey12 articles
- Journey
A year of shipping: SCULT INDIA, TechVerse Enclave, and NextUpgrad
Looking back at a year that included co-founding a studio, growing a 500-member community, and shipping five production AI apps solo.
1 min readRead - Journey
What I'd tell my first-year self about AI, college and shipping
Ten pieces of advice from the far side of an engineering degree, three hackathon wins, a failed first project and five shipped AI apps — written for the first-year who is sure everyone else knows what they are doing.
2 min readRead - Journey
Why I write: notes, blogs and building an audience that invites you
Why a GenAI engineer publishes articles and posts: to think clearly, to be found by colleges and communities, to give students something to point at — and how posts turned into sessions, judging and clients.
1 min readRead - Journey
The first 500 members: how TechVerse Enclave grew
How a college group became a 500+ member developer community with 200+ students mentored: the thirty who set the culture, the sessions that converted lurkers, and the admins who owned tracks.
2 min readRead - Journey
Leading a 14-member team while still a student
Leading engineering, design and delivery for a 14-member team at SCULT INDIA while finishing an engineering degree: written ownership, weekly demos, saying no to clients, and the mistakes of a first-time lead.
2 min readRead - Journey
What three hackathon wins (and several losses) taught me about building
Lessons from Changethon at IIT Roorkee, Product Genesis at CSJMU, BYTEBATTLE and STROT: the same team, one demo path, the honest scope slide, and how the losses shaped the judging rubric I use now.
2 min readRead - Journey
Mistakes from my first year of shipping AI products
The errors that cost real time in a year of shipping five production AI apps: no evaluation set, trusting demo data, hosting on hope, under-logging, over-promising to clients, and skipping the privacy conversation.
2 min readRead - Journey
From engineering student to GenAI engineer at an AI studio: what actually happened
The unglamorous sequence behind the title: a failed first project, paid client work with two friends, a CRM built to run ourselves, and a 14-member team — in the order it actually happened.
1 min readRead - Journey
Why I became a GenAI engineer instead of 'just' a web developer
The decision to move from building websites to building AI systems: what changed, what stayed the same, the first RAG project that made it real, and why the web skills were the foundation, not the detour.
2 min readRead - Journey
Two nights, almost no sleep: building STROT in public
48+ hours of system design, dashboards, pitch prep and coffee — what building in public during a hackathon actually looks like.
1 min readRead - Journey
From "Living Space Automation" — my first failed project — to five shipped AI apps
Every mentoring session I run starts with the same admission: my first real project failed hard. Here's why that's the part worth telling.
1 min readRead - Journey
My journey so far: Dr. Virendra Swarup, IIT Kanpur, IIT Roorkee, CSJMU, and back
A short map of the institutions and teams that shaped the last few years, from a first failed project to co-founding SCULT INDIA.
1 min readRead