floatingpoint is packaging supervised data and evals for vision AI teams

Why it matters

Vision AI companies increasingly compete on the quality of their data pipelines, not just model choice. floatingpoint is betting that pixel-level data infrastructure becomes a standalone market.

floatingpoint (@floatingpt_) launched today as an effort to build what it calls "the data layer for pixels," according to posts on X and a company overview that describes the startup as a data research lab for artificial visual intelligence.

The company says it is building off-the-shelf and custom data products for teams training, testing and shipping vision-capable models. Its stated focus areas include document intelligence, general vision, computer use, visual software engineering, spatial reasoning and long-horizon visual tasks.

The product menu is more specific than the launch slogan. floatingpoint says it offers supervised datasets for layout analysis, table structure, form understanding, structured extraction, schematic analysis, object and instance counting and other visual primitives. It also says it builds hand-crafted environments and verifiers for document reasoning, computer use, visual software engineering, game design and spatial reasoning, alongside public and private benchmarks for measuring visual capability gains in model systems.

For custom engagements, the company says it works with ML teams on schema design, distribution design, annotation guides, QA tooling, ML-assisted review, golden-set delivery, scaled production and ongoing evaluation. Its process starts with alignment on schema and edge cases, then moves through tooling, a golden set, production scale-up and delivery of data with evaluation results and learnings.

The positioning is clear: the AI data lab is going after the less visible infrastructure problem behind vision AI, where training and evaluation data is harder to structure than text. "Vision is a hard modality to build good data for," floatingpoint wrote, adding that the team "experienced this first hand" while building chunkr (@chunkrai).

So far, the company says its work has focused on helping ML teams "push the frontier of document AI," naming Extend (@ExtendHQ) as one example. It has not described pricing, customer volume, customers beyond Extend or whether it is backed by outside capital. The public funnel points to floatingpoint.ai, a scheduling page and [email protected].

The immediate ask is also hiring. floatingpoint said it is looking for engineers and researchers who care about vision, data and model behavior, and listed [email protected] as the contact. That makes today's launch both a recruiting marker and a bid to turn document-AI data work into broader vision infrastructure.