MY WORK
Projects I've shipped.
Production systems, AI tools, and research — each one built with real engineering and deployed to real users.
TLC Manufactured Home Loans — Full Platform
End-to-end lending platform for a manufactured home finance company. Consumer-facing lead capture with ZIP-code dealer matching, a dealer partner application site, and an internal CRM with KPI dashboards, Kanban pipeline, and role-based routing. All three systems production-deployed.
Amazon Prime Video — tvOS PiP
Designed and implemented tvOS Picture-in-Picture from scratch across 4 development phases. 10+ merged code reviews with zero regressions, delegate and observer pattern multi-player coordination. Authored comprehensive developer documentation covering SPM setup and platform-specific AVKit constraints.
Xcode + Bedrock Server (Amazon)
FastAPI-based OpenAI-compatible proxy enabling Amazon Bedrock model integration with Xcode 26 Apple Intelligence features. First cross-org AI integration across Prime Video, Kindle, and Q Dev teams. Docker containerized with full technical documentation.
Olympus Trading Ecosystem
Four-subsystem prediction market infrastructure in Rust. Atlas (Kalshi arb bot with RSA-PSS auth and circuit breakers), Aeolus (GEFS weather forecasting with EMOS calibration), Artemis (sharp odds aggregator with fractional Kelly sizing), and Hermes (Polymarket multi-outcome arb bot with EIP-712 signing).
Story Coder SABRAI
AI assistant pipeline leveraging OpenAI's Assistants API to classify transcripts of teachers' read-aloud sessions. Improved accuracy from 88% to 96% through few-shot prompting, Chain-of-Thought, RAG, and temperature tuning. Cross-platform Flutter front-end integrated with a FlaskAPI back-end.
Speaknolie — AI4ALL Ignite
End-to-end deepfake speech detector built in the AI4ALL Ignite accelerator. Whisper transcription + BERT + Feedforward NN achieving 81.5% test accuracy with 0.703 recall on fakes across 31,699 transcripts. Selected as 1 of 40+ teams to present at the AI4ALL recruiter symposium.
Gymnasium RL Platform
Modular reinforcement learning platform with PPO, curiosity-driven exploration via RND, and dense reward shaping. ~30% higher sample efficiency and 50% reduction in training variance on Atari benchmarks.
Madness Predictor
PyTorch neural network that predicts NCAA March Madness winners with round-by-round bracket probabilities. 70% test accuracy and 69.8% correct picks on the 2025 bracket.