A collection of notes on fine-tuning LLMs, hacking vintage hardware, and building systems that work.
Training an XGBoost model on 10 years of crash data to map the hidden geometry of risk in Austin's cycling network. A study in explainable AI and urban planning.
Investigating if safety training removes dangerous knowledge or just suppresses it. A replication of Anthropic's "Sycophancy" research on Llama-2 and Claude.
Reverse engineering a 2001 GameCube game to inject dynamic AI dialogue via memory mapping. No game code modifications required.
Building a custom pipeline to scrape, clean, and fine-tune a model on viral science videos. Improved channel Click-Through Rate from 5% to 15%.
Evaluating whether smaller open-source models exhibit the same self-awareness capabilities as frontier models like Claude 3. Testing "sycophancy" benchmarks.
Testing LLM reasoning capabilities against ancient philosophical paradoxes of vagueness. Do models understand "heaps" of sand?
Why we moved off NoSQL, the technical challenges of data migration, and how it improved our system stability. A deep dive into database architecture.
Scaling a backend overnight when a TikTok went viral. Handling traffic spikes, implementing caching with Redis, and surviving the Reddit hug of death.
My first successful open-source project. Building a CLI tool for neurodivergent developers who live in the terminal.