Projects

Selected work in computer vision, LLM systems, and full-stack development.

Attention map over both lung fields

PulmoLens

End-to-end chest X-ray diagnostic pipeline: 14 lung pathologies classified with visual explainability and guideline-grounded report generation.

Custom AttentionDenseNet (DenseNet121 + CBAM) in PyTorch with mean AUC 0.8511 and per-class thresholds calibrated to minimize false negatives; served with FastAPI on GCP Cloud Run
Gemma-powered RAG layer (LangChain + Pinecone) grounds reports in BTS, NICE, and Fleischner clinical guidelines; Grad-CAM++ heatmaps show where the model looks
Multimodal LLM eval harness with cross-family LLM-as-judge (Qwen rating Gemma) catching prompt-leakage and safety regressions in CI

PyTorch · FastAPI · LangChain · Pinecone · Gemma · Cloud Run

Say it, and it's logged

An iPhone calorie tracker where you say what you ate and it's logged. Now on the App Store.

Log by Siri, photo, barcode, or text: a language model identifies the foods and portions, then the numbers come from USDA, Health Canada, and Open Food Facts data rather than model memory
On-device estimation with Apple Intelligence where supported, with every estimate reviewable before it's saved
Apple Health sync, widgets, macro tracking, and weight trends, with no account needed to log locally

iOS · Apple Intelligence · Claude · HealthKit

Dispatch routes with live updates

Full-stack vending operations platform: procurement, dispatch, route tracking, and automated inventory reconciliation.

Next.js + TypeScript with PostgreSQL/Supabase, NextAuth RBAC, and Realtime WebSockets delivering ~50ms cross-tab dispatch updates
Supabase + Vercel wired up via MCP servers so an LLM agent can query the production DB and inspect deployments in natural language

TypeScript · Next.js · PostgreSQL · Supabase · MCP

Tracked instruments across frames

Detects and tracks surgical instruments across laparoscopic video frames for surgical workflow analysis.

YOLO11m fine-tuned on Cholec80 to detect 7 instrument classes at 83.39% mAP@0.5, with ablations validating a 5.8-point gain from augmentation
DeepSORT tracking (76.93% MOTP, 57.25% IDF1) over 3,374 frames; trajectory smoothing cut detection jitter by 36%

PyTorch · YOLO11m · DeepSORT · Python