Building local-first AI — open-weight decision models and the harnesses that make them actually useful, running entirely on your own hardware. No cloud dependency, no data leaving the machine, no per-call billing.
🔥 Featured — laya-browser-agent
A local, open-source alternative to TypeSafe Jev for browser-agent decisions. Powered by Laya, the open-source "System 1" decision model — it answers typed questions with calibrated probabilities instead of generating text, so it cannot hallucinate an instruction. Playwright/CDP drivers, MCP server, TypeSafe-compatible API included.
The fine-tuned checkpoint: ichenney/laya-browser-v32b
Frozen-encoder head fine-tune (v17→v32 recipe) that beats the official browser-tuned Laya checkpoint on 6 of 8 benchmarks — measured on identical fixtures, all eval JSONs published:
| Benchmark | 🏆 our v32b | official browser Laya | hosted Jev API |
|---|---|---|---|
| recovery2-holdout (240) | 0.7125 | 0.425 | — |
| MiniWoB (116) | 0.9138 | 0.6638 | — |
| browser-suite v4 | 0.5143 | 0.500 | — |
| browser-suite v5 | 0.5636 | 0.5818 | — |
| JevBench hard (111) | 0.4144 | 0.243 | 0.7207 |
| JevBench easy | 0.8542 | 0.979 | 1.0000 |
| decision latency (p50) | 27 ms (RTX 3080) | — | 854 ms (network) |
The honest read: in the local track (free, private, offline) it leads the
official checkpoint almost everywhere. The hosted Jev cloud API still wins on
absolute accuracy — but v32b is $0, fully private, offline, 31× faster, and
even flips score questions (0.667 vs 0.333) and temporal_numeric (0.33 vs
0.20) in its favor.
📖 Full methodology: MULTIDIM_COMPARISON.md · fresh 231-item Jev head-to-head: JEV_COMPARISON.md
| Repo | What it is |
|---|---|
| laya-browser-agent ⭐9 | The browser decision harness — element tables, guards, chunking, serve + MCP. PyPI: laya-browser-agent |
| laya-computer-use | Same decision-model approach for the desktop: reads any macOS window's Accessibility tree, a local Laya model picks the control, executes via cua-driver. No screenshots, no vision model, $0 |
| laya-training-log | Complete training log behind v32b — every version, every failure, all scripts. The repo to read if you want to reproduce the result |
| web-search | Free DuckDuckGo search + content extraction as a pip package — no API key |
| au-live-globe | 🌏 Live Australia on a 3D globe — real-time bushfires, earthquakes, flights, ships. Keyless / zero-cost (Vue + Cesium) |
| agent-skills-cn | Bilingual (EN/CN) battle-tested agent skills as open SKILL.md files |
| mcp-csv-analyst | CSV analysis over MCP |
| FluidPill | Android (Kotlin) |
Practical, opinionated workflows extracted from real use — full list:
weekly-review · web-cliplibrary · resume-localize-cn2en · project-handoff · email-deliverability-audit · expense-capture · competitor-recon · context-budget · delivery-checklist · changelog-capture · estimate-before-build · verify-claims · report-link-verification · onboarding-pack · plain-business-english · decision-records · deliverable-versioning · template-instantiator · ask-batch · backlog-triage
- Fine-tuning the computer-use model — taking
laya-computer-usefrom zero-shot Accessibility-tree decisions to a purpose-trained checkpoint (the browser fine-tuning playbook, applied to desktop UX trees) - Scaling decision-model fine-tuning beyond 322M params
- More bilingual (EN/CN) agent tooling
Open to collaboration on local-first AI, decision models, and browser/desktop automation. Find me in the issues of the repos above.