Max Caldwell
I work on small-model post-training, local inference, and agent evals. Previously CTO of a construction-AI startup and an independent AI engineering consultant. Earlier I worked at Adobe and HOVER. San Jose, CA.
GitHub|Hugging Face|LinkedIn|X
Writing
-
Fine-tuning LFM3 1.7B for Nushell tool calls (Nov 2026). 92.6% vs 22.1% and 13.0% frontier on 231 held-out tasks, 82.3% at 2.25 bits, running in-browser on WebGPU.
Projects
-
Nushell LFM3 1.7B. LoRA post-training, synthetic data generation, sandboxed WASM/WebGPU runner, and failure audit. (Demo|Model|Dataset)
-
script<T>. Scriptt is prompting with a model for types to catch agent errors early. (scriptt.ai|GitHub)
- Construction-AI Startup (CTO & Co-Founder). Architectural-drawing computer vision, 3D geometry engine, and multimodal agents (~$100K ARR, 3 engineers).