projects
Testing whether injecting live X data ahead of requests can optimize inference through KV cache reuse. Built a Rust gateway that proxies to xAI and injects shared context prefixes, with a Python benchmark suite to measure it.
Like a producer reusing loops they have made before, mainstacks lets you extract patterns from past projects and drop them into new ones. Your agents get context on how you build things, so they stop guessing and start building the way you would
Wrap your LLM client with one line of Python and every call, token, cost, and source location shows up in your dashboard in real time. Built to make agentic AI costs visible per agent, per function, and per line of code before they surprise you.
Co-authored a paper accepted to ICML 2026 AI for Science Workshop: developed a multi-agent system (MAS) that orchestrates expert decision trees with Vision-Language Models (VLMs) for automated bias labeling in forestry remote sensing, outperforming supervised ML baselines while preserving interpretability
Automates medical referral intake, OCR to structured extraction to EMR auto-fill. Taught me more about customer discovery and what people actually pay for in the context of building a startup.
My Neovim configuration. Clean, fast setup focused on development with LSP support, fuzzy finding, and Treesitter. Uses Packer.
One of my first projects, built with some friends at the first AI Hackathon from UC Berkeley. Analyzes news article bias with GPT-4 scoring and Hume AI sentiment, then rewrites the piece from three perspectives.
