I'm Deepesh — final-year CS student at IIIT Bhopal, building agentic RAG and multi-agent systems with LangGraph. I ship production-grade full-stack apps rather than leave them half-finished.
Quick facts
| 🎓 Education | B.Tech CSE, 2023 – 2027 |
| 📍 Based in | IIIT Bhopal, India 🇮🇳 |
| 💼 Experience | Full Stack Dev Intern @ Labmetrix |
| ⚔️ Competitive programming | Codeforces Specialist · CodeChef 1600+ (3★) · LeetCode Knight (~1833) |
Core stack: TypeScript · Python · C++ · React · Next.js 15 · Node.js · Express.js · FastAPI · LangChain · LangGraph · ChromaDB · LLMs · MongoDB · PostgreSQL · Docker
Break it down, prove it, ship it.
TypeScript |
React |
JavaScript |
C++ |
Python |
Docker |
Frameworks & Runtime
AI / RAG Stack
Data & Cloud
🐍 Optional add-on: the snake-eating-contributions animation needs a small GitHub Actions workflow running in a
deepu4402020/deepu4402020profile repo — happy to set that up as a separate file if you want it live.
|
Agentic RAG coding tutor, not a lookup table Stack: LangGraph | FastAPI | Next.js | ChromaDB | OpenAI
Architecture:
- 5 specialized LangGraph agents, 4 tutoring modes
- Hybrid RAG: BGE embeddings + BM25 + RRF + cross-encoder
- ~276ms avg retrieval, sub-400ms time-to-first-token
- SSE streaming, 4-level adaptive hint engine
Status: ✅ Deployed |
Agentic Chrome extension — browser automation + RAG, built to solve my own internship-search workflow Stack: TypeScript | React | Manifest V3 | LangGraph | FastAPI
Architecture:
- Dual-pipeline: tool-calling agent + structured output
- Per-tab RAG, 15K chars/page, chunked into ChromaDB
- 6 browser-action tools, 4-level DOM selector fallback
Status: ✅ Shipped |
🏋️ Eleweight — live ↗AI fitness trainer, real-time pose tracking Stack: Next.js 15 | MediaPipe Pose | WebRTC | MongoDB
Features:
- OWASP-hardened JWT auth
- Sub-100ms API responses, 95+ Lighthouse
- Solved a breaking Next.js 15 async-params bug pre-launch
Status: ✅ Live in production |
Self-written CRDT, not a library wrapper Stack: React/Quill | Custom CRDT | Stateless WebSocket relay
Highlights:
- Conflict-free concurrent edits from scratch
- Horizontally scalable relay layer
- Built in a focused two-day sprint
Status: ✅ Shipped |
|
ML security research paper Method: SVM-RBF classification, 11-class task
Result: 99.54% cross-validation accuracy
Deliverables: Paper, evaluation framework, presentation deck
Status: 📊 Research complete |
**Real-time note-taking ** Stack: MERN | Socket.io | JWT
Features: Live doc sync, rate limiting
Status: 🔧 Actively hardening (not a trophy piece — yet) |
🌍 Open Source — wger (Django/Python fitness tracker) · PR #2422, in review
Refactored WorkoutSession to use datetime_start/datetime_end instead of split date/time fields, removed the constraint blocking multiple sessions per day, added a configurable max session length, fixed timezone math feeding stats/trophies, and shipped a data migration for existing users. 18 commits · 1000+ lines changed — active back-and-forth with maintainers on session-matching priority and per-user timezone handling.
- 🌱 Interviewing for SDE / Full-Stack / AI-ML roles, campus and off-campus
- 🧠 Deepening agentic AI: hybrid retrieval, LangGraph multi-agent orchestration, eval
- 🔧 Following up on the open wger PR with maintainers
- 📄 Executive Lead at HRCC — ran the Git & GitHub hands-on workshop end-to-end
- 🎤 Building in public — technical project posts on X and LinkedIn
- 🛠️ Studying system design, cryptography fundamentals, and DevOps
+ Same instinct for DSA grinding and shipping production RAG pipelines
+ I read PR review comments closer than most people read job offers
+ 1000+ problems solved, 8+ projects shipped, 1 open-source PR in active review"First, solve the problem. Then, write the code." — John Johnson
