Hi, my name is

Saurav Chourasia.

I like to build stuff.

Computer Science student at VIT Bhopal, exploring AI, backend systems, and whatever catches my curiosity next.

See what I've built get in touch

$ whoami

saurav_chourasia

$ cat starting-point.txt

"i wonder if i can build this"

$ ls learning/

cs-fundamentals/ dsa/ backend/ agents/ ▊

0organic users on Snip & Ask
IIT-BFOSSEE Fellow '26
Resume ↓download the PDF
scroll

01. A little about me

I usually start with a question.

Sometimes it's "I wonder if I can build this." Sometimes it's "there has to be a better way to do this."

That's how most of my projects have started. I built Snip & Ask because I thought getting an answer to something on my screen should be easier. Working on an AI assistant at FOSSEE made me curious about how we can make LLMs work with information they can actually rely on. That led me into RAG, then backend engineering, and now agents and CS fundamentals.

I learn best by building things myself. The goal isn't just to make something work, but to understand why it works.

$ cat approach.md

understand what's underneath.

then build it.

$ ls interests/

retrieval/ evaluation/ backends/

B.Tech CSE '27

VIT Bhopal University · Cyber Security & Digital Forensics

CGPA 8.48 / 10
Bhopal, India

Open to internships & collaborations.

02. Things I've Built

Roughly in the order I built them. Each one exists because of what the previous one taught me.

Chrome extension · 540+ organic users

Snip & Ask

This is where it started. I wanted to ask a question about something on my screen, and the honest options were: screenshot it, save it, switch tabs, and re-explain the context. So I built a Chrome extension that captures any region of the screen and answers questions about it — without leaving the page.

The first version was just for me. Friends encouraged me to publish it, and it's now live on the Chrome Web Store. Making it work for other people was the real education: a Guest Mode with zero setup (backed by a Cloudflare Worker I run), bring-your-own-key support for Groq, Gemini, OpenAI, OpenRouter and local Ollama, an Auto Mode that routes around single-model rate limits, and on-device OCR with Tesseract.js so text-only models can read screenshots.

the arc: idea → build → make it usable → other people use it.

JavaScriptChrome MV3Cloudflare WorkersTesseract.js OCR

FOSSEE Summer Fellowship · IIT Bombay

eSim AI Assistant

At FOSSEE I worked on the eSim AI assistant, a chatbot that helps people debug circuit simulations — and it's the project that changed how I think about LLM systems.

The existing approach leaned heavily on a local LLM to reason about circuit and netlist information directly, which meant asking the model to guess facts it was never given. I redesigned the pipeline the other way around: deterministic code extracts the circuit facts first, and the LLM is handed verified facts to explain. I also built a rule-based NgSpice error analyzer that prioritizes errors by causal dependency, so root causes surface before the failures they trigger downstream.

what clicked: an LLM doesn't always need to reason itself — give it reliable information and let it explain.

PythonNgSpiceRule-based SystemsLLM Pipelines

RAG framework · built from scratch

Modular RAG Evaluation Framework

After eSim, I kept thinking about the same question in a more general form: how do we give LLMs better information? I didn't want to build another RAG demo — I wanted to understand retrieval itself, and you can't do that if you can't swap the parts.

So I built the framework from scratch — no LangChain, no LlamaIndex. Indexing, chunking, retrieval, reranking, generation and evaluation are all interchangeable modules. I used it to benchmark 4 chunking strategies against 4 retrieval pipelines on BEIR SciFact, implementing Recall@k, Precision@k, MRR and NDCG@k from scratch, and to compare generation strategies — CoVe, Cited-Span and Abstain Generation — under a RAGTruth-inspired hallucination evaluation. Everything runs locally on Ollama and ChromaDB, with cross-encoder reranking.

the point: change one component, measure, and actually see what happens.

PythonOllamaChromaDBCross-Encoder RerankingBEIR SciFact

Backend · a learning project, on purpose

LLM ChatBot Backend

This one isn't an AI breakthrough, and I won't pretend it is. After working on AI systems, I realized I didn't understand the engineering underneath them well enough. So I built a chatbot backend specifically to learn it — properly, piece by piece.

FastAPI and PostgreSQL with SQLAlchemy, JWT authentication, Alembic migrations, a layered CRUD and service structure, conversation and message persistence, Redis caching, and streaming responses over SSE — containerized with Docker and deployed. The rule I set for myself: understand the mental model first, then write the code.

the engineering under an AI app, built by hand instead of assumed.

FastAPIPostgreSQLRedisSSE StreamingDocker

03. Where I've Worked

Software Engineering Intern — eSim AI Chatbot

FOSSEE Summer Fellowship · IIT Bombay

May – Jul 2026 · Remote
  • Redesigned the chatbot's inference pipeline by extracting deterministic circuit facts before LLM inference, replacing raw netlist-to-prompt processing with structured circuit analysis.
  • Built a rule-based NgSpice error analysis engine that prioritizes errors by causal dependency, identifying root causes before downstream failures.
  • Developed a fully offline circuit analysis pipeline that generates deterministic circuit summaries without requiring an LLM.
  • Shifted the LLM's role from guessing circuit behavior to explaining verified analysis — the model no longer had to guess factual circuit information, which made the reasoning pipeline more reliable.
PythonNgSpiceLLM PipelinesRule-based Systems

B.Tech — Computer Science & Engineering

VIT Bhopal University · Cyber Security & Digital Forensics

2023 – 2027
  • CGPA 8.48 / 10 — focused coursework in data structures, DBMS, networks, and security alongside hands-on AI engineering.

04. What's Next

Right now, I'm going deeper into the fundamentals.

I'm focusing on CS fundamentals, DSA, backend engineering, and AI agents — learning how these systems work rather than rushing to the next framework.

I'm not trying to lock myself into one particular area yet. If a problem requires me to learn something new, that's part of the process.

Tools I've actually used across these projects:

Languages
PythonC++JavaScriptTypeScript
Backend
FastAPIREST APIsSQLAlchemyPostgreSQLRedisJWT Auth
AI / LLM
RAGOllamaChromaDBSentence TransformersBM25Cross-Encoder Reranking
Tools & Platforms
GitLinuxDockerCloudflare WorkersChrome MV3

05. Achievements & Certifications

  • 2026
    FOSSEE Summer Fellowship — IIT Bombay

    Selected for a nationally competitive open-source software fellowship conducted by IIT Bombay.

  • NPTEL
    Blockchain and its Applications

    Certified with Top 5% standing.

  • Live
    Snip & Ask on the Chrome Web Store

    Published extension organically reaching 540+ users.

06. Say hello

Have an idea, want to talk about something I've built, or just want to connect?

Email GitHub LinkedIn