Everything the model has shipped — course work, client platforms, and a product with users. Public repos link straight to the code.
Kuja is a live social network built around gatherings: find events near you on a map, host your own, RSVP, and meet the people who showed up.
Founder, CEO & sole engineer
The company stack, designed and driven as CTO.
Architecture, delivery & vendor management
The site you are reading: a Next.js portfolio that pulls projects live from GitHub, ships structured data for search engines and LLMs, and doubles as a printable resume.
A modular computer-vision pipeline that finds and follows small drones across thermal and RGB imagery, pairing a YOLOv12 detector with a BoT-SORT multi-object tracker and MOT-format trajectory export. Trained on a ~400K-image thermal dataset using sequence-level splits, so frames from one flight can never leak between train and validation.
Pipeline engineering, training & evaluation
A reverse-proxy firewall that secures LLM agents by separating policy-based tool access from a five-layer NLP safety pipeline: semantic scope checking, prompt-injection detection, PII redaction, attachment inspection, and multimodal tool-misuse guarding. Prompt Guard 2 and Llama Guard sit behind an LLM-assisted policy compiler with deterministic validation, so a model proposes the rule and the system still proves it.
Design, pipeline engineering & evaluation
An end-to-end ML system that scores upcoming Premier League fixtures by predicted 'liveliness', helping fans pick the best match to watch. Covers the full lifecycle: scraping and feature engineering on historical team/player stats, model iteration from R² −0.15 to 0.82, and a production web app serving live rankings.
ML engineering, modeling & deployment (team of 3)
Applied NLP techniques to three decades of English Premier League data (1993–2024) to predict team performance metrics — from preprocessing pipelines and custom train/validation/test splits to model evaluation and error analysis.
Modeling & evaluation (team of 3)
Course project building a reusable data-processing template extended across several languages, comparing model behavior and performance per language.
With Toby Xu