Cognivault Venture Labs is the data and model lab of Cognivault. We record skilled manual work from the worker's point of view across Central Asia, aggregate it into structured datasets, and fine-tune public foundation models on what we collect.
1,000 usable hours across eight domains: our first delivery target, 15 December 2026
Every recording session gains the same layers before it reaches a buyer. Pick a step, or switch the workplace.
Illustrative session. Real recordings and annotations are shared with buyers under NDA.
Trained operators wear lightweight stereo headsets during real shifts, with optional wrist cameras for hand and object detail. Our team handles configuration, charging and offloading, so workers just press record.
Every recording runs through one pipeline: ingestion, synchronization checks, quality control, task segmentation, annotation, storage and export. Originals are preserved next to processed versions.
We fine-tune public foundation models through their APIs on the datasets we build. The models learn manual tasks and make our segmentation and labeling faster with every hour collected.
Pre-seed round funded by Yammi Ventures Group, closed on 30 October 2026, to build our collection operation, data pipeline and first fine-tuned models.
Raw first-person recordings of skilled manual work collected to date.
Available collection capacity we are growing to by the end of Q1 2027.
Operations in Almaty, Astana, Bishkek and Tashkent, inside workplaces we already have access to.
Cognivault builds hyperscale AI data center infrastructure in the United States, with a 100MW project under procurement in Texas and on-site power from its own natural gas treatment. Venture Labs supplies what that compute is for: proprietary, real-world training data.
We work with robotics teams, model labs and operators who are serious about what they are building.
Contract us to collect to your spec. You define the tasks, environments, hardware and annotation schema; we run the operators, capture and quality control in workplaces we already have access to, and deliver in your format.
We record inside workplaces we already have access to across Kazakhstan, Kyrgyzstan and Uzbekistan. Available capacity grows to 100,000 hours per quarter by the end of Q1 2027, and every hour flows to Cognivault's compute in Texas.
Available capacity per quarter by the end of Q1 2027, 100,000 hours in total. Partner sites are disclosed to buyers under NDA.
Field operators by domain and city for the first collection phase.
| Task family | Almaty | Astana | Bishkek | Tashkent | Operators | Share |
|---|
First field operators by domain and city. Tashkent is ramping up.
Collection as a service: tell us the tasks, environments and formats you need, and we will contract a collection plan around them.
Our founding team are alumni of SpaceX and Agility Robotics, with 200+ projects deployed in data management and governance.
Alumni
Track record
Headcount
By the end of 2026 our team will include these roles.
Run capture hardware, charging, offloading and quality across our sites.
Work with partner sites and operators in Almaty, Astana, Bishkek and Tashkent.
Fine-tune foundation models and build model-assisted annotation.
Own the hand pose, depth and object pipelines.
Tell us who you are and which role fits. We read every application.
We will reply by email.
Book a call with the team, or send us a message and we will reply by email.
Datasets, partnerships, collection sites or press.
Opens a Google Calendar event with our team already added as a guest. Pick a time that works for you and save it to send the invite.