Building production AI systems @ Kemuri Technology
Multi-agent orchestration · Vision AI pipelines · FastAPI backends · LLM applications
AI/ML Intern at Kemuri Technology and 3rd-year AI & Data Science student. I build production-grade AI systems — not notebooks, not demos. Multi-agent orchestration, vision classification pipelines, distributed task systems, and real backends that ship and stay up.
Currently leading MAP, a multi-agent AI automation platform with LangGraph orchestration, circuit breakers, Prometheus observability, and RBAC — coordinating a team of four through code review across 8+ phases. Also building the COE-AI marine garbage classification service at Kemuri, integrating GPT-4o mini vision with production CMS pipelines.
I care about clean architecture, TDD, zero-ambiguity code review, and systems that hold up under real load.
| Domain | Stack & Scope |
|---|---|
| Multi-Agent Systems | LangGraph stateful graphs, ReAct loops, tool-calling, agent orchestration, circuit breakers |
| RAG Pipelines | FAISS/Chroma vector stores, semantic retrieval, confidence-scored decision engines |
| Vision AI | EfficientNet, GradCAM explainability, GPT-4o mini vision, OCR pipelines, deepfake detection |
| Backends & APIs | FastAPI, async Python, JWT RS256, Pydantic, SQLAlchemy, Celery + Redis, Alembic |
| Infra & Ops | Docker Compose, Nginx, Prometheus + Grafana, Neon PostgreSQL, Upstash Redis, BentoML |
| Frontend | React 18, TypeScript, Three.js, react-three-fiber, GSAP, Tailwind, WebSocket real-time |
Production-grade distributed system · LangGraph · FastAPI · Celery · PostgreSQL · Redis
Planner → Executor → Analyzer → Memory pipeline orchestrated with LangGraph stateful graphs. FAISS/Chroma RAG memory per user, ReAct executor loop with 5 real tools, confidence-scored evaluation pipeline (re-executes below 0.7 threshold), circuit breaker over all LLM calls with BentoML-served Mistral-7B fallback, RS256 JWT with RBAC, Prometheus + Grafana observability.
LangGraph LangChain FastAPI Celery Redis PostgreSQL FAISS BentoML Docker React 18 TypeScript Prometheus
Enterprise AI ticket system · 90%+ routing accuracy · FastAPI · PostgreSQL
Confidence-based decision engine auto-resolves tickets (≥ 0.75 threshold) or escalates to human agents. TF-IDF similarity search reuses solutions from resolved ticket history. Full RBAC, bcrypt auth, layered architecture with pytest coverage.
FastAPI Python SQLAlchemy NLP TF-IDF JWT PostgreSQL pytest
Local-first · Ollama · LangGraph · Docker sandboxed execution
Agent loop with tool-calling, Docker-sandboxed code execution, FAISS memory for context retrieval. Runs fully locally on qwen2.5-coder — zero cloud dependency.
Ollama LangChain FAISS Docker FastAPI React Python
EfficientNet-B4 · GradCAM explainability · FastAPI · PyTorch
Production FastAPI service serving a deepfake classifier with GradCAM visual explanations per inference. React + TypeScript frontend for real-time results.
FastAPI EfficientNet-B4 GradCAM PyTorch Explainable AI React 18 TypeScript
Campus platform · 10-phase build · FastAPI · WebSocket · Celery
Full backend for clubs, events, RSVP, waitlists, QR attendance, PDF certificate generation, budgets, real-time WebSocket notifications, and admin panel. 10 phases, full test coverage.
FastAPI PostgreSQL Redis Celery WebSocket React 18 TypeScript Neon
Odoo × VIT Pune Hackathon · built in 8 hours · OCR · Multi-tenant
Multi-tenant expense system with OCR receipt scanning, multi-step approval workflows, role-based access. Built under hackathon conditions, reached national finals.
FastAPI PostgreSQL OCR React Python Multi-tenant
AI & Agents
Backend & APIs
Frontend
Infra & Data
ML & Vision
Tooling
- Odoo × VIT Pune Hackathon Finalist — national finals May 2026, prize pool ₹45,000
- Pull Shark ×2 — PRs merged by others across team projects
- Pair Extraordinaire — co-authored commits across MAP team
- AI/ML Intern @ Kemuri Technology — production COE-AI vision service, GPT-4o mini integration
Open to collaborations on multi-agent systems, AI backends, and interesting production problems.