Automated Detection Proposes. A Person Disposes.

The studio presents each detected object as a track across the timeline, not as a series of isolated frames. A correction applies to the object's whole appearance rather than frame by frame. Merging joins tracks the detector treated as separate objects; splitting does the reverse.

An object you draw yourself is tracked forward the same way, so redacting a face that moves through a scene is one action rather than one per frame. On MP4 sources that tracking runs in the browser, so the object is followed immediately instead of waiting for a server job to return.

Correcting what the detector found

Merge duplicate tracks of the same person into one. Split a track the detector joined. Adjust a bounding box, rename an object, delete a false positive, or draw an object the detector never found and track it like a detected one.

Layers, and why they matter more than they sound

Redactions are organised into named layers, managed independently. Each carries a name, a visibility toggle, a lock, role-based permission, and its own export.

That is what makes tiered disclosure workable on a single asset. A record released to a requester, to counsel and to a court removes different things. Holding those as three layers on one asset avoids maintaining three separately redacted copies that then drift apart.

Because visibility is per layer, a reviewer sees what one release removes without disturbing another. Because permission is per layer, an analyst can work on one set while another stays locked to them. Operations apply in batch across a layer rather than region by region, and layer history is tracked.

Partial redaction

A redaction can apply to part of an object's appearance rather than all of it, so an object that needs obscuring for part of its time on screen is redacted for that span only.

Built for people who do this all day

56 keyboard shortcuts cover editing and direct assignment of 36 detection classes. The surface is fully keyboard-driven, so an analyst works without reaching for the mouse.

Work is saved as it is made, so a long session survives the browser closing or the network dropping.

• Body camera footage — use-of-force review where one person stays visible and everyone else does not
• eDiscovery — production sets with different privilege decisions per audience
• FOIA and public records — tiered release from one record
• Government — multi-department sign-off before external release

• Review is per asset in the studio. It is the considered step, not the bulk one; for high-volume routine work see automation.
• Client-side tracking applies to MP4 sources. Other formats track server-side.
• Layers organise the redaction work. The output is still a flattened rendition with the obscured regions burned in.

• Detection is automated and release is operator-gated: a reviewer corrects and adds before output is produced.
• Objects the AI does not detect are drawn manually and tracked like detected ones.
• Duplicate tracks of the same person merge into one.
• Redactions organise into independently managed layers, each with a name, visibility toggle, lock and role-based permission.
• A layer exports on its own, which is how one asset serves several disclosure tiers without maintaining separate redacted copies.
• The surface is fully keyboard-driven including direct class assignment, across 56 shortcuts and 36 classes.
• Work is saved as it goes, so it survives the browser closing.