16 year old at NCSSM. I work on agentic harness engineering, reasoning in latent space, recursive self-improvement, and synthetic data.
wizard -- self-extending autonomous coding agent in one Rust binary. Any provider (OpenAI-compatible, Anthropic, xAI) or fully local via llama.cpp. Live /evolve self-modification, MCP, messaging gateway, built-in bench. Ran its own AHE evolution loop under Grok 4.5 to drive the harness to a 100% pass rate.
agentic-harness-engineering -- contributor. Observability-driven automatic evolution of coding-agent harnesses. NexAU-AHE reaches 84.7% ± 2.1 pass@1 on Terminal-Bench 2, beating Codex, ACE, and Training-Free GRPO.
reverie -- JAX and Rust successor to Coconut. Depth-supervised halting calibrates latent compute to problem depth, so reasoning steps match hop count exactly. Includes a causal γ-ablation.
candor-bench -- honesty, sycophancy, calibration, and factuality evals for LLMs.
spore -- the most minimal Rust coding agent. One tool (bash), spawns and rewrites itself. There is a version hand-written in x86-64 assembly with no libc.
61 merged pull requests in repos I do not own, including:jax · tinygrad · Gymnasium · ratatui · zellij · rclone · petgraph · rand · maturin · insta
Rust, Python, PyTorch, JAX, SGLang, Nix, TypeScript, C++, Lua, Go. NixOS daily driver.




