Free Demo · No Signup · Air-Gapped AI
Take in every crash report, let an AI suggest, and have a person confirm every value.
Every crash report a state receives has to be checked before it becomes the official record: electronic submissions from officers' systems, scanned paper forms, and drivers' own reports. The Refinery is the one place that happens. Each report is filed into a case and read first by an AI model running on the agency's own server. A reviewer then confirms the data field by field beside the report image, while a mapper places the crash on the road network. Nothing reaches the agency's database until a person has signed off and every blocking rule passes.
The demo runs on made-up reports, with no sign-in. Pick a role to try it as: Reviewer, Location, Supervisor, FARS, or Admin. It's one shared copy that goes back to the sample data every night, so other visitors may be working the same queues. Nothing leaves the demo: no email is sent, and outside lookups use built-in stand-ins.
Before anyone opens a report, a local AI model works it over once. It checks the state's own rules and proposes a fix where a rule says what the value should be. It looks up the vehicle from its VIN, a commercial carrier from its USDOT number, and the driver's licence and registration where the state's DMV connection is set up. It reads the officer's narrative and flags coded values that don't agree with it. Every suggestion carries a confidence score and a short note on where it came from. It never changes a value on its own. The officer's value stays until a person accepts or changes it.
Air-gapped. The model runs from a file on the agency's own server, with no network connection. Crash data is never sent to an outside AI service. The only outside calls are trusted federal lookups, such as NHTSA's VIN decoder and the FMCSA carrier registry.
The Refinery is the layer before a state's crash database. It doesn't replace officers' reporting systems, and it doesn't send data to NHTSA itself. Reports come in from the systems officers already use, and the finished record goes into the database the agency already runs. From there, Crosswalk produces the national MMUCC6 format, and FARS Manifest works the fatal cases. Crash Location Resolution can feed the Location lane, and Paper Image Extraction can feed intake. Each one is a separate product, connected by a documented data contract, so an agency can adopt the whole pipeline or only the piece it's missing.
We're describing it as it is. The version described in the user manual (September 28, 2026) does all of the above today, and the public demo runs it on made-up reports. A few things aren't finished or depend on each installation:
