I am an AI engineer working on open-source medical AI: clinical NLP and de-identification for Indian languages, multimodal diagnostics, and models that can run close to the data. Recent work includes IndicPHI (a 23-language synthetic clinical corpus), Tamil medical reasoning data and evaluation harnesses, and BigTB6, a voice-driven TB and anemia screening system that won the Agentic Workflow Prize at the 2026 MedGemma Impact Challenge.

I also work on reinforcement learning for physical systems — building control, power grids, air-traffic management, and weather forecasting — and contribute fixes upstream when the tools I depend on break.

• Clinical NLP, PHI/PII de-identification, and synthetic data for low-resource and Indic languages
• Multimodal screening with small models that can run locally
• Reinforcement learning for building control, power grids, and air-traffic management
• Reproducible evaluation and open artifacts for medical AI

• IndicPHI — reproducible synthetic clinical-document pipeline for 22 Indic languages plus English. It generates PHI/PII-annotated records, audits them with an LLM judge and deterministic validators, curates with NeMo Curator, and exports GLiNER-ready train/eval splits. Python · NeMo · GLiNER

• tamil-med-sft — pipeline that designs, generates, judges, and decontaminates native-Tamil medical reasoning SFT data. Sarvam-105B generates items, Grok-4.3 audits them across six dimensions, and salvageable revisions are re-judged and promoted. Python · SFT · data curation

• TamilMedicalMILU and TamilMedicalMMLU-Pro — standalone, uv-managed lm-eval runners for Tamil health and medicine benchmarks, scored against any OpenAI-compatible vLLM endpoint. lm-eval · vLLM

• BigTB6 — voice-first screening for TB and anemia: cough audio, chest X-ray, palm/eye/nail pallor, and webcam rPPG respiratory rate, orchestrated with Gemini Live and MedGemma. Agentic Workflow Prize, MedGemma Impact Challenge 2026. Gemini Live · MedGemma · multimodal

• Anemia screening models — MedSigLIP linear probes and QLoRA fine-tunes for palm, eye, and fingernail anemia detection, with weights released on Hugging Face. MedSigLIP · QLoRA

• google/sbsim — upstream issue reports and fix PRs for Google's Smart Buildings control simulator: SAC actor crash, baseline schedule-policy action order, and local comfort-window handling. RL · building control

• grid2op-openenv — OpenEnv-compatible IEEE 14-bus grid environment with four topology-control tasks and a Think → Simulate → Act LLM baseline; the server owns simulation state. OpenEnv · Grid2Op

• Complete_RL_Journey — Connect Four AlphaZero from scratch: game environment, MCTS, ResNet policy/value network, parallel self-play, and a released checkpoint. AlphaZero · PyTorch

• MedGemma Impact Challenge — Agentic Workflow Prize (2nd) · Google Research × Kaggle · 2026
  For BigTB6, a voice-driven TB and anemia screening demo.

• Bio x AI Hackathon — Winner · bio.xyz · 2025