SAI

Python, Anthropic API, GitHub Actions, GitHub Pages // 2026

• ›Built an autonomous trading system that scrapes SEC filings and Congressional disclosures, scores insider conviction using a custom weighted model (stake increase, role seniority, recency decay), and places real trades on a demo account twice weekly with zero human intervention.
• ›Integrated Claude Opus as the portfolio manager, feeding enriched candidate data (fundamentals, RSI, price history, news) to generate up to 5 picks per run with allocation percentages, confidence scores, and written rationale referencing specific insider signals.

PyTorch, RunPod, HuggingFace // 2025

• ›Pretrained GPT-2 Small from scratch on a curated 18B token dataset assembled from 7 sources (FineWeb-EDU, arXiv, RefinedWeb, Gutenberg) across 8 x A100 GPUs, building a custom memory-mapped dataloader to handle large-scale data streaming beyond RAM capacity.
• ›Instruction-finetuned the pretrained model on 17M tokens across multiple SFT datasets, outperforming the original GPT-2 Small on TruthfulQA and MMLU benchmarks and deployed an interactive demo on HuggingFace Spaces for public evaluation.

Python, FastAPI, Docker, PostgreSQL // 2025

• ›Architected and shipped a fully on-premise AI platform for a security company — 5 production apps including a private LLM assistant, Agentic RAG, ASR transcription, document OCR, and an AI coding agent, eliminating all external LLM API dependencies on sensitive data.
• ›Built an internal model evaluation framework benchmarking 10+ open-weight models across reasoning, instruction following, code quality, and summarization to drive continuous data-driven model selection.

Python, OpenCV, MediaPipe, PyTorch, Three.js // 2024

• ›Built a computer vision pipeline for Ajax that detects and scores footballer shooting technique from match footage, extracting body keypoints and computing biomechanical metrics across the kinematic chain.
• ›Rendered a 3D skeleton overlay on detected shots using Three.js, giving coaches an interactive frame-by-frame breakdown of technique vs. ideal form.

PyTorch, GAN, Bayesian Deep Learning // 2023

• ›Master's thesis: incorporated uncertainty estimation into a GAN-based super-resolution model (ESRGAN), enabling the network to output calibrated confidence maps alongside upscaled images.
• ›Applied to medical imaging — allowing clinicians to see not just the reconstructed image but how confident the model is in each region, improving diagnostic trustworthiness.