Laya Decision Studio Laya Decision Studio

Overview

Built a local FastAPI workbench for typed yes/no, choice and score decisions over text, images and video frames, with an evaluation harness that separates working software from model quality.

Highlights

1. 01

   Served the Laya decision model and Qwen3-VL-2B vision locally with pinned model revisions, explicit device reporting and bounded image and timestamped-frame inputs, plus labeled evaluation that reports accuracy, Brier score, NLL and ECE.

2. 02

   Ran an independent 24-request Coding Lab benchmark in which Laya got more individual decisions right than a rules baseline (238/288 vs. 226/288) but fewer complete specifications (2/24 vs. 5/24), and reported that result as measured.

3. 03

   Connected the model to Codex through an MCP server and CLI with a versioned repair ledger, and kept it out of review filtering after a held-out diagnostic caught only 1 of 12 positive cases.

Stack

• Python
• FastAPI
• PyTorch
• Hugging Face Transformers
• Qwen3-VL-2B
• SmolVLM2
• Laya
• Model Context Protocol (MCP)
• Pytest
• Model Calibration
• CUDA