█████╗ ███╗   ██╗██╗    ██╗ █████╗ ██╗   ██╗
██╔══██╗████╗  ██║██║    ██║██╔══██╗╚██╗ ██╔╝
███████║██╔██╗ ██║██║ █╗ ██║███████║ ╚████╔╝ 
██╔══██║██║╚██╗██║██║███╗██║██╔══██║  ╚██╔╝  
██║  ██║██║ ╚████║╚███╔███╔╝██║  ██║   ██║   
╚═╝  ╚═╝╚═╝  ╚═══╝ ╚══╝╚══╝ ╚═╝  ╚═╝   ╚═╝  

I'm a B.Tech CSE student at IIIT Surat who builds AI-integrated, production-grade systems — not just demos.

• 🔬 Deep interest in ML systems, NLP pipelines, and LLM application development
• 🏗 I build real tools — job automation bots, collaborative editors, scraping engines
• 🌐 Full-stack capable: from REST APIs and WebSockets to React frontends
• 🔁 20+ open-source contributions with 70%+ PR merge rate
• 🎯 Looking for: SWE / AI-ML Engineer roles (internship or full-time)

Built a miniature version of vLLM to understand production LLM serving at the systems level. No model.generate() used anywhere.

┌─────────────────────────────────────────────────────────┐
│                   docker-compose stack                   │
│                                                          │
│   React Dashboard ──REST──▶ FastAPI :8000                │
│   • GPU memory gauge        POST /generate (SSE)         │
│   • TPS sparkline           GET  /stats                  │
│   • Batch slot viz          WS   /ws/stats               │
│                                    │                     │
│                             Inference Engine             │
│                             • Custom token loop          │
│                             • Continuous batching        │
│                             • Temperature + top-p        │
│                                    │                     │
│                          GPT-2 / TinyLlama weights       │
└─────────────────────────────────────────────────────────┘

What I built:

• Custom generation loop — tokenize → forward pass → logits → temperature scaling → top-p nucleus sampling → decode, fully hand-written in PyTorch
• Continuous batching scheduler — when any sequence hits <EOS>, its slot is immediately freed and the next queued request fills it without pausing other active sequences (3–5x more GPU-efficient than naive batching)
• FastAPI + SSE streaming — tokens stream to the client one-by-one as generated, exactly like ChatGPT's interface
• React live metrics dashboard — GPU memory gauge, tokens/sec sparkline, queue depth, batch slot visualizer, all updating every second Stack: Python PyTorch FastAPI React Vite Docker CUDA HuggingFace Transformers WebSocket SSE

Hackathon-winning real-time collaborative coding platform with AI-powered code suggestions

• Built room-based live collaboration using Django Channels + WebSockets
• Integrated Hugging Face inference API for context-aware code suggestions
• Handles concurrent users with real-time sync and conflict-free editing Stack: Django WebSockets JavaScript Hugging Face API

Languages

AI / ML

Web & Backend

Tools & DevOps

• ✅ Contributed to 20+ repositories across ML, DevTools, and Web categories
• 🔀 70%+ PR merge rate — focused on meaningful, well-scoped contributions
• 🐛 Contributions include: bug fixes, feature additions, docs improvements, test coverage

• 📚 Deepening expertise in transformer architectures and fine-tuning workflows
• 🔧 Building more LLM-powered developer tools
• 🤝 Open to collaborations, internships, and full-time SWE/ML roles