Projects
autoresearch-win-rtx
Windows-first fork of Karpathy's autoresearch for autonomous single-GPU LLM pretraining on consumer RTX hardware. Agents iterate on `train.py` under a fixed 5-minute budget, keep/discard experiments by `val_bpb`, and use runtime autotuning for VRAM-tier-aware batch sizing and checkpointing. This project is listed as a notable fork in the upstream repository.
ShipSpec Claude Code Plugin
A Claude Code marketplace plugin that brings spec-driven development workflows into the CLI. It packages planning and implementation commands (`/feature-planning`, `/implement-task`, `/implement-feature`), specialized agents, reusable skills, and Stop-event hook scripts that enforce retry loops and task verification against PRD/SDD artifacts.
segov.dev
An AI-powered Next.js portfolio centered on a terminal-style AMA built with AI SDK Agents. It retrieves private context from Vercel Blob, records production traces in Neon, and uses those traces to fine-tune Qwen3.5-4B, served scale-to-zero on Modal with vLLM.