Lens. Video, indexed by meaning. Conversation, attached to the frame.
A video intelligence surface built on the same ultra-smooth playback substrate that powers Observatory’s session review. Frame-level tagging, structured store-and-retrieval, semantic search across a library, and an LLM you can hold a conversation with — about your video, with the moment it cites still on screen.
Tag. Store. Search. Talk.
Tag. Frame-level and span-level tagging that survives playback. Built-in detectors for speaker, scene, action, on-screen text, and chart. Custom taggers in your own model class plug in as a tool call.
Store. A purpose-built media surface designed for hours-long footage and dense annotation. Tagged interfaces for retrieval — by speaker, span, scene, transcript phrase, or visual signature. We hold the index; you hold the policy.
Search.Semantic and structured search across a library: “every moment a parent asks about safety,” “every clip where the whiteboard shows X,” “every silence longer than four seconds.” Results are returned as cited spans, not transcripts.
Talk. An LLM conversation attached to a video. Ask a question; the model answers with the moment cited and the player parked on the frame. Pairs with Helm for a coach to steer the conversation in real time.
Three shapes, one engagement model.
Lens is offered as enterprise engagements only. Pricing is part of the contract; volume curves, storage tiers, and committed-use discounts are posted on request.
- Sized for small archives
- Default tagger library
- Standard playback latency
- Observatory traces included
- Higher volumes, dedicated capacity
- Custom taggers & rubrics
- LLM-over-video conversations
- SSO, RBAC, exports, audit log
- Private deployment, your region
- Indefinite retention with your policy
- Bring-your-own taggers, confidential
- On-call review of indexing protocols
Why playback matters as much as the model.
Most video AI products treat playback as a thumbnail problem. We treat it as the product. We built a frame-accurate, low-latency scrubbing engine for Observatory that holds up at hours of footage and thousands of overlaid annotations. Lens sits on that same engine — so when the LLM cites a moment, the player is there before you finish reading.
That is the difference between an LLM that talks about a video and an LLM that reads it with you. The model is only as useful as the moment it can put on the screen.
Stop transcribing video. Start having conversations with it.
Lens is offered as enterprise engagements only. Tell us about the footage — the volume, the use, the constraints — and we will tell you whether Lens is the right shape.