https://pranjulrathour.github.io/ 2026-10-02T09:35:00Z Pranjul Rathourpranjulrathour41@gmail.com https://pranjulrathour.github.io/what-is-machine-learning/2026-09-24T08:00:00Z

Machine learning is software that improves at a task from data instead of following rules a programmer wrote by hand. Supervised, unsupervised and reinforcement learning explained plainly, how ML relates to AI and deep learning, and where it is genuinely used today, from spam filters to the models behind modern generative AI.

https://pranjulrathour.github.io/what-is-a-vector-database/2026-09-24T08:00:00Z

A vector database stores embeddings and finds the nearest ones fast. How vector search and ANN indexes work, when a library or Postgres is enough, and how to choose between FAISS, pgvector, Qdrant, Chroma, Weaviate, Pinecone and Milvus.

https://pranjulrathour.github.io/what-is-prompt-engineering/2026-09-24T08:00:00Z

Prompt engineering is writing and testing the instructions a language model runs on. The techniques that hold up, how system prompts differ from user prompts, what prompt injection is and how to defend against it, how to learn it, and where it sits next to RAG, agents and fine-tuning.

https://pranjulrathour.github.io/what-is-an-llm/2026-09-24T08:00:00Z

A large language model predicts the next token, and everything else follows from that. How LLMs are trained, what parameters, tokens and context windows mean, how GPT, Claude, Gemini, Llama and Mistral differ, what hallucination is, how to run one locally, and how to reason about cost. Written from shipping LLM products.

https://pranjulrathour.github.io/what-is-vibe-coding/2026-09-24T08:00:00Z

Vibe coding is building software by describing what you want to an AI assistant and accepting what it produces. What the term means, how coding agents differ from autocomplete, which tools students should learn, the projects it suits and the ones it wrecks, the review habits that keep it safe, and whether it makes learning to code pointless. From shipping production systems with and without assistants.

https://pranjulrathour.github.io/what-is-computer-vision/2026-09-24T08:00:00Z

Computer vision is the field of getting machines to extract meaning from images and video: what is in the picture, where it is, and what is happening. How it evolved from hand-engineered features to convolutional networks to vision transformers, the core tasks (classification, detection, segmentation, tracking), where it is genuinely used today, and how it relates to the vision-language models that now sit alongside it.

https://pranjulrathour.github.io/ai-final-year-project-ideas/2026-09-24T08:00:00Z

Final year project ideas in AI that a student team can actually finish and demo: RAG over college documents, document extraction with validation, face-recognition attendance with a privacy boundary, receipt and prescription readers, a fine-tuned small model, speech-to-text workspaces, and more. For each: what it is, the hard part, the tech, and what makes it stand out. Plus how to pick, how to scope for a semester, what evaluators look for, and how to turn the project into a portfolio piece. From building Annapurna, FaceVision, DocuLens and RAG.NextUpgrad.

https://pranjulrathour.github.io/bca-vs-btech-for-ai-career/2026-09-24T08:00:00Z

The honest comparison of BCA and B.Tech CSE for students who want AI and GenAI careers in India: curriculum depth, maths, placements, eligibility for roles and higher studies, cost, and the gaps each degree leaves. Plus the plan that closes those gaps, from a BCA graduate who shipped five production AI systems.

https://pranjulrathour.github.io/best-free-ai-coding-assistants-for-developers/2026-09-24T08:00:00Z

Which AI coding assistants have a real free tier, what each is actually good at, autocomplete versus agentic editing versus terminal-based coding agents, and how to choose without paying for a tool a free tier already covers. Compared: GitHub Copilot's free tier, Claude Code, Cursor, Windsurf/Codeium, Gemini Code Assist, and open alternatives, from daily use across production codebases.

https://pranjulrathour.github.io/data-analyst-vs-ai-ml-engineer-career-path/2026-09-24T08:00:00Z

Three career paths students often confuse because they share tools: data analyst, machine learning engineer, and AI or GenAI engineer. What each role does day to day, the real skill overlap and where it ends, typical entry points from a BCA, B.Tech or MCA background, and how to decide between them without guessing from job titles alone.

https://pranjulrathour.github.io/dev-to-vs-hashnode-vs-medium-for-developers/2026-09-24T08:00:00Z

A practical comparison for developers deciding where to publish technical writing: Dev.to's community and free tier, Hashnode's custom domain and now-paid API, Medium's paywall and algorithm, and a self-hosted blog you fully own. What each costs, who reads each, SEO and discoverability differences, and the two-platform strategy that works for most people: write once, publish to your own site first, syndicate the rest. From publishing across all of them.

https://pranjulrathour.github.io/do-you-need-a-gpu-for-ai-vram-and-hardware-guide/2026-09-24T08:00:00Z

Most AI work a student does does not need a personal GPU at all. What actually needs local GPU memory versus what an API or a free cloud notebook already covers, how much VRAM common tasks need, quantization's effect on fitting a model into less memory, and a plain decision guide before spending money on hardware or a cloud GPU rental.

https://pranjulrathour.github.io/fastapi-for-ai-applications/2026-09-24T08:00:00Z

Why FastAPI is the default choice for serving LLM and RAG applications in Python: async request handling that does not block on slow model calls, dependency injection for clean provider and database wiring, background tasks for long-running jobs, streaming responses for token-by-token output, and the request lifecycle mistakes that quietly break concurrency. From the FastAPI layer behind a production RAG platform and a document extraction service.

https://pranjulrathour.github.io/free-ai-tools-for-students/2026-09-24T08:00:00Z

The AI tools that actually change how a computer science student learns and builds: coding assistants, free model access, notebooks with free GPUs, local model runners, vector databases, deployment platforms, and 15 free browser tools plus a 1,211-prompt library I built. What each is for, its free-tier limits, and how to use it without letting it think for you.

https://pranjulrathour.github.io/from-bca-student-to-genai-engineer/2026-09-24T08:00:00Z

The honest path from a BCA at VSICS Kanpur to shipping five production AI systems and leading engineering for a 14-member team: the community I built, the hackathons I won and lost, the first clients, the products, and the decisions that compounded. Written for students at colleges nobody has heard of.

https://pranjulrathour.github.io/genai-interview-questions/2026-09-24T08:00:00Z

The questions GenAI engineer interviews actually ask in 2026 and how to answer them from experience: LLM fundamentals, RAG, fine-tuning, prompt engineering, agents, evaluation, cost and system design, plus behavioural questions about projects and failures. From interviewing candidates at SCULT INDIA and being interviewed.

https://pranjulrathour.github.io/hackathon-judging-criteria-and-rubric/2026-09-24T08:00:00Z

The four criteria most hackathon rubrics reduce to, with weights, score anchors and the questions that reveal each; a 15-minute judging protocol; how to brief judges and handle conflicts of interest; tie-breaking; feedback that helps teams; and how to become a judge. From judging and briefing panels at college hackathons and winning three.

https://pranjulrathour.github.io/hackathon-pitch-deck-template/2026-09-24T08:00:00Z

A hackathon presentation that fits in three minutes: seven slides in the order judges want, what goes on each, a word-for-word script skeleton, how to demo without a crash, how to answer Q&A, and the mistakes that lose otherwise strong teams. Includes the Smart India Hackathon adaptation. From a builder with three first prizes who also judges.

https://pranjulrathour.github.io/hackathon-project-ideas/2026-09-24T08:00:00Z

Forty hackathon project ideas across AI, agriculture, education, health, campus life and fintech, each with the user, the demo and the hard technical part named, plus how to pick one, how to adapt it to a problem statement, and what judges reward. From a builder with three first prizes and time on the judging panel.

https://pranjulrathour.github.io/how-does-face-recognition-work/2026-09-24T08:00:00Z

Face detection finds faces; face recognition turns each face into an embedding and compares distances against a threshold; liveness detection stops a photo from passing. How each step works, why systems fail on twins and low light, how accurate they are, the privacy rules, and how I built a browser-only version that never uploads a face.

https://pranjulrathour.github.io/how-does-ocr-work/2026-09-24T08:00:00Z

OCR turns images of text into text. How the pipeline works, where Tesseract still fits, how vision-language models changed document AI, how to OCR a PDF for free, how to extract invoice fields you can trust, and how I built a page-batched OCR and speech workspace on Mistral.

https://pranjulrathour.github.io/how-does-speech-to-text-work/2026-09-24T08:00:00Z

How automatic speech recognition turns audio into text, what changed with Whisper, how to build real-time transcription with streaming and voice activity detection, how to handle Hindi and mixed-language speech, the accuracy metric that matters, and how I built a live speech-to-text feature into a document workspace.

https://pranjulrathour.github.io/how-much-does-an-llm-api-cost/2026-09-24T08:00:00Z

LLM APIs charge per million tokens, input and output priced separately. How to estimate the monthly cost of a feature before you build it, why RAG prompts are expensive, what prompt caching, batching, routing and smaller models save, when self-hosting an open model is cheaper, and the caps and alerts that stop a viral day from becoming a bill. From running five production AI apps on tight budgets.

https://pranjulrathour.github.io/how-to-become-an-ai-engineer/2026-09-24T08:00:00Z

The skills a GenAI or AI engineer actually uses, a roadmap that works from any college including BCA and tier-3, the projects that get interviews, how to think about salary and remote roles in India, and what I did between a Kanpur classroom and shipping five production AI systems.

https://pranjulrathour.github.io/how-to-build-a-developer-portfolio-website/2026-09-24T08:00:00Z

A practical guide to building a personal website as a developer or CS student: the pages it needs, the project write-up format that gets read, what to leave off, free hosting and a custom domain, performance and accessibility basics, and how to keep it updated without it becoming a second job. From building and rebuilding a personal site that anchors a speaking and hiring brand.

https://pranjulrathour.github.io/how-to-build-an-ai-app/2026-09-24T08:00:00Z

The architecture of an LLM app, a Python and FastAPI stack that ships in a weekend, free hosting options that still work in 2026, how to keep API costs predictable, the security and logging every AI app needs, and the mistakes that kill student projects after demo day. From shipping five production AI apps in six months.

https://pranjulrathour.github.io/how-to-evaluate-a-rag-system/2026-09-24T08:00:00Z

Evaluate retrieval and generation separately. How to build a labelled question set from real usage, measure recall at k and MRR for retrieval, faithfulness and answer relevance for generation, use RAGAS and LLM judges without fooling yourself, run evaluation in CI, and read the numbers that predict whether users will trust the system. From a production RAG platform with 196 tests.

https://pranjulrathour.github.io/how-to-fine-tune-an-llm/2026-09-24T08:00:00Z

What fine-tuning an LLM means, when it beats RAG or prompting, how LoRA and QLoRA make it fit on a free GPU, how to prepare a dataset, read the loss curve and evaluate honestly.

https://pranjulrathour.github.io/how-to-get-an-ai-internship/2026-09-24T08:00:00Z

A practical route to a first AI or GenAI internship for students at any college: which companies actually hire interns for AI work, how to build the three projects that get replies, how to write a cold email founders answer, how to use hackathons and communities as the pipeline, remote and freelance options, and how to turn an internship into a job. From someone who found work without a placement cell.

https://pranjulrathour.github.io/how-to-give-a-guest-lecture-in-college/2026-09-24T08:00:00Z

How to get invited to speak at a college, how to prepare a 60- or 90-minute technical session students remember, how to structure slides and demos, how to handle a room that goes quiet or a question you cannot answer, and what turns one talk into repeat invitations. From speaking and mentoring at colleges in Kanpur and running a 500-member developer community.

https://pranjulrathour.github.io/how-to-improve-public-speaking-for-tech-talks/2026-09-24T08:00:00Z

Practical public speaking improvement for engineers and students giving technical talks: why content-first practice beats generic public-speaking advice, how to structure a technical talk so a room follows it, handling nerves and filler words, reading a room in real time, and the specific practice routine that improves a talk faster than general public speaking courses.

https://pranjulrathour.github.io/how-to-invite-a-guest-speaker-to-your-college/2026-09-24T08:00:00Z

A checklist for inviting a guest speaker or guest lecturer to a college event: how to find one, what to put in the first email, whether speakers are paid, what a speaker needs from organisers, how to introduce them, and how to turn one session into a lasting mentor relationship. Written by someone who receives these invitations.

https://pranjulrathour.github.io/how-to-learn-python-for-ai/2026-09-24T08:00:00Z

Python is the language of AI, and most students learn the wrong parts of it. The exact Python an AI engineer uses, a twelve-week roadmap from zero to a deployed LLM application, the libraries that matter, what to skip, free resources, and the mistakes that keep students in tutorial loops. From shipping five production AI systems in Python.

https://pranjulrathour.github.io/how-to-negotiate-your-first-tech-job-offer/2026-09-24T08:00:00Z

Most freshers accept the first number because they assume freshers cannot negotiate. What is actually negotiable in an entry-level tech offer in India, how to ask without a competing offer, the exact wording that works, and why almost anyone can negotiate at least once without risking the offer.

https://pranjulrathour.github.io/how-to-participate-in-your-first-hackathon/2026-09-24T08:00:00Z

A beginner's guide to hackathons for students: where hackathons are listed (Devfolio, Unstop, MLH, Devpost, Hack2skill, Smart India Hackathon), how to register and get shortlisted, how to find teammates when you know nobody, what to install and prepare the week before, what the 24 to 48 hours actually feel like, what beginners should build, and how to make the first one count even if you do not win. From a first hackathon that went badly and three that were won later.

https://pranjulrathour.github.io/how-to-prepare-a-dataset-for-fine-tuning/2026-09-24T08:00:00Z

The dataset is the fine-tune. How to write a spec, choose between Alpaca, ShareGPT and chat formats, apply the model's chat template correctly, decide how many examples you need, generate and filter synthetic data, deduplicate, split, and validate every row before spending a GPU hour. From building a fine-tuning platform.

https://pranjulrathour.github.io/how-to-start-a-tech-community-on-campus/2026-09-24T08:00:00Z

How to start and grow a student developer community or coding club: the first session, the weekly rhythm that matters more than promotion, roles and leadership handover, getting speakers and sponsors, running hackathons, and what TechVerse Enclave's growth to 500+ members and 200+ mentored students actually looked like.

https://pranjulrathour.github.io/how-to-structure-an-llm-project-in-python/2026-09-24T08:00:00Z

The project structure for a Python LLM application that stays maintainable after the first demo: a src layout with clear modules for providers, prompts, retrieval and evaluation, typed settings from environment variables, prompts stored as versioned files, structured outputs with Pydantic, a golden test set and an eval command, logging that makes failures debuggable, and packaging for deployment. From the layouts used in a production RAG platform and a document extraction system.

https://pranjulrathour.github.io/how-to-win-a-hackathon/2026-09-24T08:00:00Z

What a hackathon is, how to find and join one in India, how to pick a problem, build in 24 to 48 hours, pitch in three minutes and demo without a crash. Written from three first prizes and the judge's table.

https://pranjulrathour.github.io/how-to-write-a-tech-resume-for-freshers-in-india/2026-09-24T08:00:00Z

A practical resume guide for CS, BCA and MCA freshers applying for AI, software and engineering roles in India: the one-page structure that survives both an ATS and a human reader, projects over certificates, how to phrase them as outcomes, what recruiters actually search for, and why a simple resume website often beats a PDF for the roles that matter most.

https://pranjulrathour.github.io/hybrid-search-and-reranking-in-rag/2026-09-24T08:00:00Z

Why vector search alone misses exact terms and keyword search alone misses meaning, how to fuse them with reciprocal rank fusion in a dozen lines, what a cross-encoder reranker does and why it beats retrieving more, how to tune k at each stage, how to measure the gain with recall, and how the hybrid path works in a production RAG platform.

https://pranjulrathour.github.io/is-ai-a-good-career-in-india/2026-09-24T08:00:00Z

An honest answer for students and career-switchers asking whether AI is a good career in India: which roles exist (GenAI engineer, ML engineer, data scientist, AI product engineer, applied researcher), how hiring actually works for freshers, what companies test for, whether you need a degree from a top college, how remote and startup paths differ from campus placements, the realistic risks, and a plan for the first eighteen months. From a BCA student in Kanpur who became a GenAI engineer without a placement cell.

https://pranjulrathour.github.io/lora-vs-qlora/2026-09-24T08:00:00Z

The practical differences between LoRA, QLoRA and full fine-tuning: how much VRAM each needs for 1B to 70B models, how fast they train, how quality compares, which to choose for your task and hardware, the rank, alpha and target-module settings that matter, and a free-GPU recipe. From building a fine-tuning platform that runs QLoRA from the browser.

https://pranjulrathour.github.io/mentoring-students-remotely-and-internationally/2026-09-24T08:00:00Z

What a good technical mentor does differently from a teacher or a manager, how to find one when your college has no formal program, how to ask well, and how mentoring works once it crosses a time zone: async-first communication, recorded sessions, and running a student mentorship program for a community that spans countries. From mentoring 200-plus students in Kanpur and speaking to international student communities remotely.

https://pranjulrathour.github.io/mini-ai-project-ideas-for-weekend-builds/2026-09-24T08:00:00Z

Not every AI project needs a semester. Fifteen weekend-scale ideas, a receipt splitter, a WhatsApp study-group summariser, a plant identifier, each scoped to two days, with the one hard part and the API or model that makes it possible. For students who want to ship something small and real between bigger projects, and build the habit of finishing.

https://pranjulrathour.github.io/personal-branding-for-developers/2026-09-24T08:00:00Z

How a student or early-career engineer builds a reputation that produces internships, clients and speaking invitations: what to post and where, the LinkedIn profile and GitHub that recruiters actually read, writing technical posts without a following, turning projects into talks, and the automation and consistency that make it sustainable. From growing a community to 500+ members and speaking at colleges.

https://pranjulrathour.github.io/rag-vs-fine-tuning/2026-09-24T08:00:00Z

RAG supplies knowledge at answer time; fine-tuning changes behaviour; prompting steers both. When each wins, when to combine them, what each costs to build and maintain, the failure modes, and the decision tree I use with clients after building both a RAG platform and a fine-tuning platform in the same year.

https://pranjulrathour.github.io/redis-and-caching-strategies-for-llm-apps/2026-09-24T08:00:00Z

A cache hit on an LLM call is nearly free; a miss costs real money and real seconds. Exact-match caching with Redis, semantic caching for near-duplicate questions, what to cache safely and what never to cache, cache invalidation for RAG when source documents change, and the latency and cost numbers that make caching worth building before you need it.

https://pranjulrathour.github.io/smart-india-hackathon-guide/2026-09-24T08:00:00Z

What Smart India Hackathon is, how registration works through your institute, how to read and choose a problem statement, what the internal round eliminates most teams on, how to build the prototype and the presentation ministry judges expect, and what happens after. From a builder with three hackathon first prizes and time as a judge.

https://pranjulrathour.github.io/system-design-for-ai-applications/2026-09-24T08:00:00Z

AI system design interviews ask you to design a RAG chatbot, a document-processing pipeline, or an agentic system under real constraints: latency, cost, scale, failure modes. The framework: clarify requirements, sketch the pipeline, name the hard trade-offs, then go deep on the two or three components that actually matter. Worked through a RAG system and an agent example, with the failure modes interviewers are listening for.