Selected Projects

A few projects that represent how I like to design, build, and ship things.

Physical.AI

Co-Founder · SF

Data-driven marketplace for physical ads. Built a Zillow-style discovery experience on a map; achieved top 6 at YC W25 hackathon finals out of 100+ teams.

  • Integrated live attention score using Caltrans PeMS traffic data to estimate OTS volume, dwell time, and impact by time of day
  • Backend ingests ad placements, maps to nearby sensors, computes metrics against historical baselines, and surfaces ticker-style scores

Sympli

Technical Co-Founder

Voice-first health companion app. Built voice intake → clinician-style note pipeline (ASR + summarization) with latency under 1 minute.

  • Deployed HIPAA-aware AWS stack (TypeScript + Postgres + self-hosted LLM)
  • Shipped alpha to 20 users (100+ sessions) with end-to-end instrumentation

TAP-Score (Push-T OOD Detection)

Research · PyTorch

Replicated Diffusion Policy on Push-T; built TAP-Score to rate (obs, action-chunk) pairs using demo positives + shuffled/noisy negatives.

  • Tested OOD detection under brightness/noise/occlusion perturbations
  • Reporting AUROC and clean false-alarm rate with score traces

1:72 Scale HVAC Assembly Part Counter

Computer Vision · Contract

Freelance CV project to count small parts in 1:72 HVAC assembly kits from video and image data. Delivered an iOS app for real-time part counting.

  • Trained object detector across 14 part classes; achieved 95% accuracy on a 1,200-image test set
  • Improved robustness to lighting variance and motion blur; optimised latency to under 5s on iPhone 16 Pro

Transformer LM from Scratch

PyTorch · Personal

Implemented a Transformer language model from scratch in PyTorch; trained on an anonymised corpus (2.6M tokens).

  • Reached validation perplexity of 15
  • Full implementation of attention, positional encoding, and training loop

NLP Final Project: Legal MCQA

NLP · BERT · CaseHOLD

NYU NLP project on legal multiple-choice QA over the CaseHOLD dataset, using BERT-based models and a new evaluation metric combining difficulty, confidence, and correctness.

  • Compared BERT-Double, Legal-BERT, and Custom-Legal BERT models
  • Implemented end-to-end NLP pipeline over legal case corpora
View project

Crypto Signal Dashboard

AI x Crypto · TypeScript

Real-time dashboard for monitoring high-priority crypto trading and risk signals. 2–3k lines backend TS, 1k lines frontend React with sub-2s latency.

  • Defined key signals across on-chain and exchange data
  • Designed for quick, actionable decision-making with streaming aggregation
View project
View all repositories on GitHub