An AI research & product lab
We build multimodal and agentic AI: systems that connect what they see, hear, and read to what they do next.
A product-photography example.
Watch the edit. Understand the checks. Then try it.
The left view is the original. We’ll edit the right view into a warm studio shot. The headphones must keep their blue finish, shape, and size.
Why this matters: a convincing AI edit can still break the brief. Check what changed and what must stay the same. Original synthetic scene with supplied geometry; scripted guided edits and computed scene checks.
Our applications bring AI into personal and team workflows. Our developer tools address the details that make those workflows dependable: identity, code quality, and evaluation.
A personal AI that sees,
listens, and remembers.
Talk naturally, recall captured messages, read with your camera, and return to saved details. Hindi-first, accessible, and built for continuity.
Shared context.
Coordinated work.
Connect company information and agent workflows through operational state: what changed, what is current, and what can happen next.
From a name
to the right entity.
Connect references to your catalog, carry corrections through a conversation, and ask for clarification when an action needs a more specific answer.
Know what worked.
Understand what failed.
Evaluate RAG, agents, and document workflows. Inspect tool calls, compare runs, and examine the evidence behind a score.
Code review you can
trace back to the change.
Review pull requests in Codex or Claude Code. Inspect finding evidence, see coverage gaps, and resume large reviews. Optional Multivon scoring uses human-adjudicated results.
How does an observation become a dependable decision? We study visual representations, memory, and agent behavior—and build experiments that make their failures visible.
Every agent says “posted”. The evidence tells six different stories.
The final balance looks correct, but the audit trail retains the forbidden changes.
1. Post → 2. Change → 3. Restore → 4. Finish
Follow the selected execution through persisted state and protected-record history. The charts use recorded observations; their vertical scales adapt to each series. Acceptance requires both checks. Missing evidence blocks acceptance.
Reproduce all six runs ↗The scene supplies a start, a goal, and occupied cells. Change the scene to give the agent a different problem.
Runs in your browser: a synthetic scene, deterministic planner, and outcome checks. A working illustration of the loop, not a neural-model benchmark.
Computer vision and multimodal learning connect pixels, language, and time. Video and 3D representations preserve structure across views; world models connect state to possible outcomes.
Representations & world modelsAgents need to know what a request refers to and which facts still apply. Entity resolution grounds those references; memory carries context forward. Planning then connects that state to a permitted action.
Context & agent systemsCheck the final state and the path taken to reach it. Separate perception errors from planning failures, compare baselines, and measure the cost of getting the task right.
Evaluation & learningThat feedback loop is already concrete in Multivon. Our published TAT-QA study found that asking for answers and evidence together reduced answer accuracy and increased estimated model cost. A useful result is one that changes an engineering decision.
Read the studyOne model and protocol, using supplied table and text contexts. This study does not measure visual-agent or customer performance. Protocol, per-question metrics, and limitations are public.
Research, products, partnerships, investment—or just curiosity. We’re always happy to hear from you.
hello@omnitensorlabs.com