I build and evaluate production ML & AI systems

Gundata: Agentic AI

A traditional Andhra Pradesh Sankranti dice game rebuilt as a testbed for AI agents — from rule-based bots to a trained reinforcement learning opponent, with live explainable decision-making.

• Tabular Q-learning agent over an 18-state MDP, trained via 20,000 self-play episodes
• Explainability dashboard logging each agent's reasoning and top-3 Q-values per move
• Clean MVC game engine with two rule variants and multiple AI opponent personalities

MediTrust

Explainable cardiovascular risk prediction — ROC-AUC 0.87, per-patient SHAP explanations, full-stack FastAPI + React + PostgreSQL deployed on AWS.

• Logistic Regression model achieving ROC-AUC 0.87 on a held-out test set
• Per-patient SHAP explanations so every prediction includes a feature-contribution breakdown for clinical trust
• Full-stack production system: FastAPI backend, React UI, PostgreSQL persistence, Dockerized services

LLM Evaluation · Handshake AI

Comparative evaluation of frontier LLMs in an agentic coding environment — custom task design, structured scoring, and detailed rationales.

Campus-Objects

LW-DETR multi-class detection on 3,000+ annotated campus images — 0.71 mAP@0.5, COCO-format pipeline.

• Built and annotated a 3,000+ image campus dataset in COCO format with consistent labeling guidelines
• Fine-tuned LW-DETR via transfer learning, reaching 0.71 mAP@0.5 across multiple object classes
• Applied augmentation and evaluation pipelines for reproducible detection benchmarking

Snap Tune

Multi-modal music recommender — BLIP image captioning + GPT mood inference + Spotify retrieval.

• BLIP image captioning for semantic scene understanding from user-uploaded photos
• DistilGPT2 mood inference bridging vision output to music-relevant emotional context
• Spotify API retrieval and playlist generation, deployed as an interactive Streamlit Cloud app