I build AI systems that are grounded, measured, and shipped — not just notebooks. My work centers on retrieval-augmented generation, multi-agent orchestration, and the evaluation discipline that separates a demo from something you can trust in production.
- 🎓 Studying AI/ML at IIT Guwahati, with a BSc background
- 🔭 Currently building Daedalus — a grounded RAG platform for AI/ML interview prep
- 🔬 Running VESPER — an evidence-first exoplanet detection pipeline on TESS/Kepler data
- 🌱 Going deeper on LLM evaluation, vector retrieval, and systems engineering
- 💬 Ask me about RAG pipelines, hybrid retrieval, agent orchestration, or MLOps
- 🎯 Open to Applied AI / ML Engineer roles in 2026
LLM & retrieval · Hugging Face Transformers · sentence-transformers · ChromaDB · BM25 + RRF · Ollama · Groq · MCP · spaCy
Data & analysis · NumPy · pandas · Matplotlib · Seaborn
Engineering · Pydantic · SQLAlchemy · uv · pytest · Ruff · mypy · pre-commit
| Project | What it does | Stack | Highlight |
|---|---|---|---|
VESPER |
Evidence-first exoplanet transit detection over TESS/Kepler light curves — detects single events, infers periods from spacing, and reserves full-spectrum search for stars without local evidence | Python · NumPy · MAST/TESS SPOC · pytest | Phase I sealed and audited: recall non-inferior to full TLS search (−0.48 pp, inside a −2 pp margin) across 3 statistical methods. Phase II underway |
Daedalus |
AI interview-prep platform running grounded RAG over your own study material, with a dedicated evaluation module | FastAPI · sentence-transformers · Transformers · PyMuPDF4LLM · PaddleOCR · uv | Building the full RAG path — ingestion → embeddings → retrieval → generation → eval — typed with Pydantic and gated by ruff/mypy/pre-commit |
TK-Shield |
Detects bio-piracy by semantically matching documented traditional knowledge against patent claims, then drafts citation-backed opposition documents | FastAPI · ChromaDB · BM25 · sentence-transformers · spaCy · Ollama · React + TS | 100% precision@1 and @5 on landmark cases with zero false positives; hybrid retrieval via Reciprocal Rank Fusion over 16,292 real US patents |
ReliefAI · live ↗ |
Multi-agent disaster relief orchestrator — one natural-language command dispatches five specialist agents for scheduling, supply chain, and multilingual comms | Groq (Llama 3.3 70B) · FastMCP · FastAPI · SQLAlchemy · Docker | 5 coordinated agents, 6+ languages, deployed on Cloud Run at zero infrastructure cost |
Also worth a look: mlops-env — a 3-task RL environment for production MLOps built on OpenEnv · ai-from-scratch — ML fundamentals implemented from first principles · Pulsar Prediction — imbalanced classification with SMOTE + XGBoost
learning: [ LLM evaluation, vector retrieval at scale, systems engineering ]
building: Daedalus — grounded RAG with a real evaluation harness
researching: VESPER Phase II — triage impossibility theory & monotransit detection
goal: Land an Applied AI Engineer role and ship one RAG system people rely onI'm looking for Applied AI / ML Engineer opportunities — if you're hiring or just want to talk about RAG and agents, reach out.
