Visai

Visai is a self-improving agent that makes a model faster on the machine it actually runs on, then remembers what worked so the next run searches less. You give it a model, a workload, and a quality budget. It measures the stock model, finds the layers that cost the most time, and has an agent write faster kernels for them. A candidate is kept only if it matches the model’s own layer on hidden inputs and makes the whole model faster in paired runs without leaving the quality budget. It then searches deployment settings on top of those kernels — quantization, chunking, streaming, post-processing — and reports unoptimized against optimized. Every trial, lesson, failed idea, and winning kernel is stored in MongoDB Atlas and retrieved on the next run. It targets Apple Silicon with MLX and Metal. The models it has run are Qwen3-0.6B, Parakeet for speech recognition, and Nemotron for speaker diarization. On an M3 Pro that produced 1.67× faster decode for Qwen, 1.17× faster transcription for Parakeet at the same word error rate, and diarization error cut from 35.6% to 15.5%. Built with MongoDB Atlas, OpenRouter, and the Strands Agents SDK.