View project →Activation-Weighted Low-Rank Compression
Research paper proposing an activation-weighted fitting method that reduces perplexity degradation by 5-7x vs plain SVD across GPT-2, Gemma-3, and Qwen2.5.
Projects
Curated project work across model training, product interfaces, and systems experiments.
View project →
View project →Velora Custom LM Trainer
A from-scratch PyTorch trainer for a compact LLaMA-style language model with resumable training and guided structured generation.
View project →AI Fitness Trainer
A privacy-first, locally-run AI fitness coach combining YOLOv11 pose estimation with a vision-language model for real-time rep counting, form analysis, and AI coaching.
View project →CityBuilder — A Self-Contained Micro-City Simulation with Local AI
A full-stack micro-city MVP with a FastAPI simulation engine, React/TypeScript dashboard, interactive canvas map, and a local Q-learning policy agent — no external AI APIs required.
View project →MagneticStorage — Biologically-Inspired Agentic Memory
A local, biologically-inspired agentic memory system that organizes information into hierarchical concept groups and cells, going beyond flat RAG with three-stage hybrid search, entity routing, and LLM-driven consolidation.
View project →Forest Protector
A Python tower-defense game with 10 escalating waves, 50 dynamic path combinations, tower upgrades, economy, and a live GitHub Pages build.
View project →AI Face Detect and Recognition WebGPU
A client-side AI face tracking and recognition application running YOLOv11 face detection and EdgeFace embeddings directly in the browser via WebGPU and ONNX Runtime.