The engine inside.
Vision and OCR built for legal evidence — degraded scans, handwriting, night footage. Licensed under your brand, running in your building or a sealed single tenant.
Two evidence floods, one bottleneck.
Deed books, medical records and handwritten field notes come back technically processed and practically unsearchable. Generic OCR was built for clean invoices. Footage is worse — most of it is still cleared by hand.
Unsearchable documents
The market rate to process one gigabyte for review — with the hardest pages, the ones that decide matters, coming back empty.
Un-redactable video
Manual redaction labor per footage-hour. A majority of agencies still do it by hand: thirty-plus labor-hours to clear one or two hours of film.
Statutes do not wait
California's deadline for releasing critical-incident footage. New York allows 48 hours — and agencies cannot bill requesters for the redaction labor.
Motorola / CDG 2022 · ComplexDiscovery 2024 · Veritone 2026 · Cal. AB 748
One engine, two pipelines.
You integrate once. Both lanes emit the same findings schema, and every finding carries provenance — page and line for documents, frame and second for footage.
Reads what the commodity APIs return blank.
- Degraded-scan and handwriting OCR
- Layout and table extraction, structure preserved
- Bates validation across productions
- Redaction detection — what was covered, and where
- Every output cited to page and line
Clears footage without a room full of contractors.
- Scene segmentation and per-frame metadata
- Face and plate detection with automated redaction
- Audio redaction on the same timeline
- Transcription and speaker attribution
- Every finding cited to the second
Sealed. Input and output are the contract. One API, one findings schema, under your brand — we never appear in front of your customer.
May the evidence travel?
That is the only question that decides the architecture, and it is your customer's answer, not ours. Both paths run the same engine and return the same schema.
The building
An appliance on your customer's own hardware. Nothing leaves the premises — no egress, no third-party API, no telemetry. Privileged and sealed material never crosses a network boundary.
- Sited
- Firm or agency hardware
- Egress
- Zero
- Billed
- Licensed capacity
The cloud
A single-tenant instance, metered by the hour. No shared inference, no pooled storage, no training on your data. Source media is purged to zero once findings are returned.
- Sited
- Isolated single tenant
- Egress
- Findings only
- Billed
- Metered hourly
Built by litigators, for the record.
Most vision systems are built to be impressive. This one was built to survive a hearing on its own output — first in real matters, then productized. That difference shows up in three places, and each is something your customer's opposing counsel will ask about.
Pinned model revisions and hashed weights. The same input produces byte-identical output, so a result can be reproduced months later — by you, or by the other side.
An immutable audit log from ingest to finding: what ran, which version, on which bytes, when. The record is an artifact of the run, not a report written afterwards.
Sensitive stages are gated on an explicit, logged, per-matter authorization. Absent it, the stage refuses to run — the default is silence, never a quiet inference.
A twelve-month gap, closed in ninety days.
Hiring computer-vision engineers takes three to six months if the requisition fills, and legal-grade edge cases take longer than that. Two senior salaries are at risk before a customer ever touches the feature. Relativity, Reveal, Exterro, OpenText and Motorola all bought their engines.
| Dimension | Commodity APIs | Doctrine Engine |
|---|---|---|
| Target input | Clean, born-digital | Degraded scans, handwriting, night footage |
| Bates & redaction detection | Absent | Native, with an audit trail |
| Video evidence | Generic object labels | Redaction plus searchable metadata |
| Reproducibility | Models change under you | Pinned versions, byte-identical reruns |
| Deployment | Their cloud only | Your building, or a sealed single tenant |
| Whose brand | Theirs | Yours |
Your hardest data. Seven days. Zero cost.
We do not ask you to believe a benchmark. You set the metrics before we start, your team grades the output, and you run it side by side against Textract, Google and Azure on your own material.
Intake
Five hundred pages, or ten hours of footage — the material your current stack handles worst. Metrics agreed in writing first.
Scored results
Output returned against the metrics you defined, graded by your people, next to the commodity APIs on the same input.
Decision
License it, or walk away. Pilot data is deleted within fourteen days. One pilot per partner — there is never a second.
Wholesale pricing is per gigabyte for documents and per footage-hour for video, tiered by volume against an annual minimum. Rate card on request under NDA.
Give us your hardest documents.
Scored results in seven days, against metrics you define, graded by your team. No cost and no obligation — if we miss your numbers, you have lost one email.