I build production-grade LLM applications, RAG pipelines, and multi-agent systems — from architecture to deployment. Shipped FoLiOai, a live AI SaaS for Amazon India sellers.
Production deployments, research-grade ML, and systems built from scratch.
AI-powered financial intelligence platform for Amazon India sellers. Deterministic GST/TCS/TDS reconciliation engine parses settlement reports, detects leakage across 5 categories, and produces per-SKU P&L with source-row audit trails. Gemini is isolated strictly to narrative generation — all figures come from deterministic code. Shipped with Firebase Auth, BullMQ async queues, paywall, and DPDP Act 2023 compliance.
Multi-stage research pipeline: question decomposition → DuckDuckGo/Wikipedia retrieval → HTML extraction → ChromaDB semantic search → LLM synthesis with contradiction detection and source traceability. Multi-provider LLM abstraction (Groq, Gemini, Anthropic), retry logic, SQLite persistence, and 30+ unit/integration tests.
End-to-end GNN pipeline on 1,128 ESOL compounds. Graph Isomorphism Network achieved MAE 0.634 ± 0.035, R² 0.811 — 29% lower MAE than Morgan-MLP baseline across a 5-seed controlled study with random and scaffold splits. Streamlit inference app with 2D molecular visualisation, PubChem lookup, and Tanimoto applicability-domain analysis.
Real-time Indian Sign Language recognition supporting 27+ gestures. MediaPipe handles hand landmark extraction; Groq/Llama 3.1 powers contextual tutoring feedback. Owned the full pipeline from landmark extraction through LLM feedback generation, containerised with Docker for production at SuprMentr.
PyTorch-style dynamic computation graph built from scratch in pure Python. Operator overloading and reverse-mode automatic differentiation. Trained MLPs from scratch on Iris and Moons datasets — demonstrates first-principles understanding of how modern frameworks work under the hood.
Tools I've used to ship things.
What I've built at each place.
Looking for AI Engineer and GenAI Engineer roles in Bengaluru, remote, or hybrid. Particularly interested in RAG systems, multi-agent architectures, and production LLM applications.