Laya replaces LLM-as-a-judge with a 322M decision engine — 26,639 stars in 9 days
AIFrontierReads · reddit · 2026-09-28
- Laya turns decisions — routing, triage, yes/no calls — into typed outputs from a 322M model instead of generated text, with a routing-only CLI, triage presets, and an abstention gate below a confidence threshold.
- Hands-on: the author ran the full tutorial on CPU, including French ticket classification. With minconfidence=0.90, Laya abstained on one case it would otherwise have misclassified — the honest highlight. Warm latency: 0.7s per question on CPU.
- Caveat: checkpoints tend to be over-confident (calibration issue), so don't trust raw scores blindly. The project hit 26,639 GitHub stars in 9 days.
More from coding & agent
- "Plan mode was cool in 2025": devs poke fun at Google's Antigravity — BLUECOW009 · 2026-09-28
- GraphMemix: open-source graph memory lifts multimodal agent memory accuracy by up to 11.75 points — TheTuringPost · 2026-09-28
- Building a phone agent that catches missed callbacks: setup, costs, and what broke — Purple_Lab5333 · 2026-09-28
- Vory: open-source iPhone app for self-hosted Hermes agents, now in public beta — Matt0975 · 2026-09-28
- One-shot Claude prompt generates a scroll-driven comic-timeline news site — thomas_unise · 2026-09-28
- Agent swarm simulator replays task DAGs to cut expensive swarm experiments — KyeGomezB · 2026-09-28