Ambit

A hardware startup builds forty units of a prototype. Optical inspection machines start around fifty thousand dollars and have to be programmed per board, which never pays back at that volume, and there is no defect data to train on because the board did not exist last month. So small teams inspect by eye, or skip it. Quality control is one of the reasons hardware iterates slower than software. Ambit makes inspection cost fourteen photos and three seconds. You photograph known-good units and it fits a specialist for that exact part from normal examples only, with zero defect labels. Show it any part afterwards and it embeds the frame, vector searches a registry of specialists, and loads the one trained on the closest imagery. When nothing in the registry has competence over what it is looking at, it refuses instead of passing an uninspected board. That refusal is the hard part. Five specialists are registered and a sixth board is withheld. Top-1 routing is 125/125. Per-model coverage gates refuse 25/25 out-of-registry frames, while every global threshold we tested refuses 0/25, because the withheld board scores 0.9399, above all of them. It genuinely is a PCB and the registry is full of PCBs. Only a per-model gate declines it. The cost, stated plainly, is that 6 of 125 in-registry frames are refused too. MongoDB Atlas is the system, not a store beside it. Routing is $vectorSearch over model_router_idx, one 512-d OpenCLIP centroid per model. Weights are PatchCore coresets in GridFS at 7.55 MB each. PatchCore has no gradient-trained parameters: training stores a coreset of patch features and inference is nearest-neighbour against it, so a model is remembered examples of normal. Vector search picks which memory applies, nearest-neighbour searches inside it. Findings carry their own embedding indexed for episodic recall, LangGraph checkpoints live in the same cluster so a kill -9 mid-training resumes instead of relearning, and fleet health runs as aggregation pipelines where the data already is. LangGraph runs the agent graph with node-level checkpointing and cold-start as a branch. Fireworks runs Qwen3-VL to describe the difference between the defect crop and the same region of the golden reference. OpenRouter does structured-output adjudication inside the 0.04 ambiguous band with model fallbacks. ElevenLabs is a voice analyst over five MongoDB-backed tools including explain_refusal and recall_similar, and every claim it makes appears as a tool call first. The demo is live. A phone streams a viewfinder to the projected view, holds still, and uploads one full-resolution frame. An Arduino board is cold-started on site from fourteen photos, comes back nominal untouched, and is flagged with a localized heatmap and a written description once a pin is bent. Built during the event: the Arduino specialist and local ingest path, Fireworks VLM defect narration, OpenRouter adjudication, ElevenLabs voice, project grouping in the registry, the capture and analysis interface, and the trends view. The README separates this from pre-existing work in full.