Did it actually follow the procedure?
A controlled process either ran inside its procedure's limits or it did not, and the limits are written down. Miriel holds the video, the machine telemetry and the governing document on one timeline, so that question has an answer with evidence attached.
// the document, the telemetry, and the footage — joined, not adjacent
This is the compliance problem where video provably is not enough.
Most "compliance" use cases are vague. This one is not: there is a document, a measurable process and usually a regulator. Parameter ranges, preconditions, material handling, operator qualification — all written down, all checkable.
What breaks is that the operation is frequently unwatchable. A welding arc saturates the sensor. A machining centre runs with the door shut and coolant on the glass. A part cures inside an oven. A cold-chain pallet is a sealed box. And the operator is occluded behind a hood, a respirator or a machine enclosure at exactly the moment you most need to know who ran the step.
Even with a clean view, "conforms to WPS-104" is not a visual class. No detector emits it. It is a join between a video interval, a telemetry series and a PDF. The compliance signal comes from the machine, the context comes from the cameras, the rule comes from the document — and something has to hold all three on one timeline and answer across them.
So we built the thing that holds all three.
- The document corpus gets in properly first. Drawings, scans, stamped certs, redlines and tables are where procedure limits actually live, and naive text extraction loses them. The parser pulls structured elements — titles, narrative, tables as tables — and falls back to page-image rendering with OCR when text extraction is thin. Page images and figures are embedded into CLIP space and indexed alongside the text, so a drawing is retrievable by what a vision model reads off it, not only by whatever OCR caught.
- Multi-document joins, resolved at write time. "Which procedure covers this operation on this material at this thickness, for this client's spec revision C, and is the person assigned currently qualified?" is a four-document join, and embedding similarity is not transitive — no value of k retrieves those chunks together. Entity and relationship extraction runs on ingest, and queries fan out across vector search, metadata filtering and a depth-2 graph traversal in parallel, then merge and re-rank.
- Counts are counts, not samples. "How many parts in this package need heat treatment?" cannot be answered by top-k retrieval — it returns k chunks, and the answer needs all of them. An explicit exhaustive mode decomposes aggregation-shaped questions map-reduce style: a structured sub-query against every relevant document, then an aggregation over the results.
- Telemetry on the same timeline as the video. Process parameters, temperatures and material handling land in one time series beside the detections and transcripts, written in batches with idempotency keys and retried rather than dropped.
- Evidence packets, per operation. An exception names the interval, the frames either side, the clause it breached, who ran it and whether they were qualified — assembled when it happens and still queryable a year later, which is the form an auditor actually wants.
- “Did the welds on job 4418 run inside the WPS amperage range?”
- “Which controlled operations last week have no evidence packet?”
- “Was anyone at that station without a hood down?”
- “How long did that assembly sit waiting for fit-up?”
- “Who was the last person to handle that spool of 316L, and where did it go?”
Every answer comes back with its evidence — the clip, the row, the clause — so it can be checked rather than trusted.
Measured on a real customer corpus. The economics move deliberately from expensive read-time inference to cheap write-time extraction, which is also what makes the joins expressible at all.
Compliance monitoring is only as good as the corpus underneath it, so the document work gates everything downstream — we would want to settle a retrieval bake-off with measured recall@k and MRR before instrumenting a single cell. On the floor side, every facility is its own calibration problem: camera placement, object vocabulary, what telemetry a PLC or historian already records, and what is allowed to leave the building.
GET STARTED
Bring us a question you cannot answer today.
The fastest way to find out whether this fits is to pick one real question your current stack cannot answer, and see what it takes.
// no spam, just a beta waitlist