FactoryLens · Trusted answers over plant-floor video
Ask your cameras anything. Act only on answers that prove themselves.
FactoryLens sits between a vision model and your EHS team. Every answer must cite the frames it relies on, agree with what is already known about the scene, or say it cannot tell, before anyone acts on it.
8 recorded model answers, graded in the open: 5 cleared, 1 held for review, 1 not actionable, 1 abstained. See every answer
Pick a clip
Sewing line0:10· Garment line, close shot of three stations
An operations event: the far-left station is unmanned in every frame while the others work.
Ask the footage
Instant answers replay a recorded model answer or read the clip’s event record, at no cost. Live questions use one call from a small daily allowance.
Pick a question. Every answer arrives with the gate's verdict on top, and each timestamp plays the moment it cites.
What the model sees: 10 stills
These are the exact stills the model is sent, cut once from the stock clip with identifiable faces masked. The player shows the source footage. Live questions go to a third-party vision model (openai/gpt-4o) through OpenRouter under the same routing rules as uploads.
Question log
Every answer shown on this page is logged with its source and the gate’s verdict. Each entry’s SHA-256 covers the one before it, so editing a stored entry shows up as a broken chain on reload. It lives in this browser only and has no outside anchor, so it cannot stop someone who recomputes every hash; a deployment anchors the chain server-side in write-once storage.
Nothing logged yet. Ask the footage a question and the entry appears here.
Stored in this browser only; nothing here is sent anywhere.
The one-page answer
The problem
Asking video a question in plain language is now standard. Verkada, Spot AI, Coram, Protex and Intenseye all offer it. But a vision model answers a question about a forklift near-miss just as fluently when it is wrong, and the person who acts on that answer carries the liability.
What FactoryLens does
Holds the model to a typed safety and operations ontology, forces every answer to cite the stills it used, and runs a deterministic gate that holds an answer that contradicts the labels it is checked against, invents an event or cites a moment it never saw, before the answer reaches a supervisor. Every question is logged.
The proof so far
8 unedited answers from the live model on labelled sample clips, recorded 26 Sep 2026. The gate cleared 5, held 1 for review against the human labels, and marked 1 not actionable for citing no evidence; the model abstained on 1. That demonstrates the gate; it is not an accuracy estimate, which is measured on a buyer’s own cameras.
Where it stands
Running here: face masking before any frame leaves the browser, the live model, the event record, the gate and the question log. Still design-stage: the on-site edge appliance, camera ingestion, role-based access and multi-site rollup.