PROJECTS
Selected research and production builds
Immuny — Health Ally
Clinical AI allergy tracker in production, routing between AWS Nova Micro and MedGemma 7B (4-bit quantized), exported to ONNX for edge inference.
Tech Canary — Automated Content Posting
Automated content-posting pipeline running in production for a private client; scheduling and publishing handled end-to-end.
PMNet — Verizon
3D autoencoder for RF pathloss prediction — 98% fidelity across 19K+ geospatial samples, 57% MAE reduction.
LLM Benchmarking — Google
10 frontier LLMs evaluated; findings shaped model-selection calls.
School Facilities from Imagery
Four facility attributes measured for 25 US public schools from overhead imagery, with a calibrated confidence on every field.
Nika — MoE Transformer from Scratch
2×1M-param Mixture-of-Experts transformer built from scratch (RoPE, SwiGLU, RMSNorm) with KV-cached inference, DDP + hand-rolled ring all-reduce, and an Apple Metal port.
Intelligent Bistro
Conversational AI food-ordering app — order by voice, text, or tap. 236 tests at 94% coverage, strict TypeScript, CI/CD pipeline.
RL Autonomous Driving
PPO agent on a custom ray-casting car simulation; reward shaping fixed a reward-hacking failure and raised mean survival from 28 to 510 steps over 100k timesteps.
Insurance Policy Intelligence Engine
Upload an insurance policy and ask questions in plain English — browser-side PDF parsing and a Python Cloudflare Worker that answers with cited excerpts.
Immune Repertoire Profiling from RNA-seq
TCR/BCR repertoire reconstruction from bulk RNA-seq with MiXCR and TRUST4 — clonal diversity, V/J usage and CDR3 analysis across 7 immune loci.
RLVR Fine-Tuning — Qwen2.5-7B
Fine-tuned Qwen2.5-7B with GRPO on the tau2-bench benchmark to output structured tool calls instead of prose.
RepoRaven — Paper/Repo Reproducibility Auditor
Agentic auditor that reads an ML paper against its own code, flagging data leakage and hyperparameter mismatches with the exact line and the sentence it contradicts.
Language-Editable Spatial Staging
Stage objects inside an imported 3D room by voice or text — an LLM emits spatial relations, a deterministic resolver turns them into transforms, and a vision model validates each placement.
Appraisal Document Parsing with Voice
Parses long-form appraisal PDFs into structured summaries and reads them back through a Whisper-based speech pipeline.
Sentiment-Driven Portfolio Optimizer
Black-Litterman optimization + FinBERT sentiment signals. Sharpe 1.43, alpha 6.2%, beats passive S&P 500.