Contrivance Labs
The research arm. Where the models get trained, evaluated against real material, and made small enough to run on hardware we own.
- Multimodal understanding
Joint models over video frames, speech and on-screen text, built so that a query about a place or a moment can be answered from what was actually recorded.
- Self-hosted inference
Serving open-weight models on our own GPUs rather than renting capacity, and measuring what that costs per user.
- Retrieval over real footage
Search that reads the video, not just its caption.
- Evaluation
Blind, corpus-grounded comparisons before anything changes what a model sees.
