Compliance by TwelveLabs - Video compliance review powered by rules you control
by•
Compliance by TwelveLabs is a SaaS application that reviews your video library against rules your team writes, not ours. It ingests footage, applies your own compliance rule packs, and returns reviewer-ready findings with context, not just a timestamp and a label. Powered by TwelveLabs' Pegasus model, it explains why a moment may violate a rule so reviewers can accept, reject, or annotate in one queue.
Replies
TwelveLabs
I’m Simon, Head of Field Engineering at TwelveLabs, part of the team behind Compliance by TwelveLabs.
The problem we kept coming back to: if reviewers have to rewatch the entire video to validate AI flags, what have we actually saved them?
Compliance review is still largely manual, and every market brings different rules. Detecting “violence” isn’t enough. Reviewers need to understand what happened, in what context, and why it matters under the policy they’re applying.
That’s what we built this around:
Context reviewers can act on. Pegasus explains findings instead of just labeling them.
Rules compliance teams own. Adapt regional packs, edit rules, tune thresholds, and publish versions without waiting on us.
Synthetic media detection built in. NVIDIA’s Synthetic Video Detector adds frame-level scoring alongside our contextual analysis.
Findings ready for review, with signed reports and API access.
We’re targeting a reviewer-rejection rate of 15% or less. The goal is straightforward: less time chasing false positives, more time on decisions that need human judgment.
What’s an edge case you’d want to put this through? A scene that’s acceptable in one market but restricted in another? Something AI consistently flags incorrectly? I’d love to hear it and learn where we still have work to do!
@simon_lecointe Having the findings ready for review with API access makes integrating it sound a lot smoother, congrats for shipping
@simon_lecointe Having the important context and blink ready should make the review process much less boring, congrats the team!
Adjust Page Brightness - Smart Control
@simon_lecointe explaining the reason behind a flag instead of just dumping timestamps saves so much time.
The reviewer queue makes a lot sense. Nobody wants to rewatch 20 mins just to verify one flag.
How does it handle rules that depend more on context than what's actually visible?
TwelveLabs
@evan_taft1Great question, it's not just about what's visible in the video. A lot of rules depend on where the content will run, which broadcaster it's going to, the audience, the time slot, etc. The same ad could be fine for one broadcaster or region and need changes for another.
So the idea is to combine what we understand from the video (visuals, dialogue, text, overall context) with the information provided for the review. An ad airing during children's programming might have different restrictions than the same ad airing late at night. Different countries have different requirements, and even two broadcasters in the same country can have their own standards around language, alcohol, disclaimers, etc.
We shouldn't expect the model to guess where the ad is going or who it's targeting. That context needs to be provided so we can apply the right rules. If something is missing or open to interpretation, we should flag it, explain what's missing, and leave it for human review instead of forcing a pass/fail.
The different market angle is pretty interesting. I can see that being useful for larger content teams.
Kilo Code
humbled by this new collaboration with @c3lim and the @TwelveLabs team.
Fun fact: this is their 6th launch on @Product Hunt, their 5th in the last 10 months. Behind the scenes, there's a formula we started shaping when we launched Marengo 3.0 and Pegasus 1.5, and it's all about consistency.
Learn more in this thread in /p/twelvelabs.
Launch, and keep launching!
TwelveLabs
@hamza_afzal_butt It's easy to adapt, the app’s rule evaluation framework lets us add country-specific requirements and test them against real videos. We can check what gets flagged, understand why, and refine rules that are too strict or miss something. So it's mainly about configuring and validating the local rules, not changing the app itself.
Voquill
Interesting use of AI and this could save reviewers a ton of time. Congrats team!
This could be interesting for teams where compliance checks are still mostly manual. I’d love to see how accurate the explanations are on real world footage.
@edward_paul3 Yeah, that’s probably the part we’re most focused on getting right.
Happy to share some real examples. Is there a specific type of content or violation you’d be interested in seeing?
I appreciate the focus on reviewer-ready results rather than just dumping AI detections on the user. The human review part feels well condsidered.
Dial
targeting a 15% reviewer-rejection rate makes sense as a false-positive metric, but for a compliance tool I'd actually worry more about the other direction - the violation it never flags at all, so nobody ever puts eyes on that clip. false positives get caught by a human in the queue eventually. a false negative just quietly never shows up. is there a way to sample the "cleared" footage to catch what Pegasus missed, or does the whole system rely on the rule pack being complete upfront
@galdayan Good catch, and I agree. That’s probably the harder failure mode for a compliance tool.
Today, the coverage is really based on how complete the rule pack is. Pegasus will apply the rules you give it, but if something is missing from the rule pack, it can obviously miss it.
One thing we’re looking at is spot checking some of the content that was cleared, as another QA layer, to help identify gaps or drift in the rule pack itself.
Would be interested to hear what you would want to see from something like that.
Dial
@lior_berezinski_tl random sampling of cleared clips would catch some of it, but I'd weight the sample toward the clips where Pegasus's own confidence was closest to the decision boundary rather than sampling uniformly. a clip it cleared with high confidence is probably fine; a clip that just barely cleared is exactly where a gap in the rule pack would hide. the other thing I'd want is for whatever the spot check finds to feed back into the rule pack automatically as a suggested addition, not just a flagged miss someone has to manually translate into a new rule - otherwise the QA layer catches the same category of gap over and over instead of closing it.