Figure Eight Federal and the intelligence before AI

A radar image does not look like a photograph. A laser scanner produces a cloud of points. Neither arrives with a helpful little note explaining what matters. Before an AI model can learn from either, somebody must make the world legible to it. That is the territory Figure Eight Federal occupies: the space between collecting information and teaching a machine what to do with it.

F8F, as the company calls itself, belongs to Appen. Its business combines software with people who understand the material being labeled. The product is partly a dataset, partly a workflow, and partly a way of checking whether the instruction a machine received was worth giving.

01 / The sensor needs a translator

Consider Artemis, F8F’s enterprise labeling platform. The company describes native support for synthetic aperture radar, LiDAR and multispectral imagery. Native matters here. A defense team may need to label the original sensor data, rather than squeeze it into a format designed for ordinary pictures. The distinction sounds technical until the format becomes an obstacle to the job.

Artemis sits alongside Athena, an NLP tagging solution, and Gen AI Optimizer. F8F’s services extend to data labeling, model fine-tuning, sentiment optimization, data fusion and LLM evaluation. Together, these offerings address a familiar purchasing problem: a team can possess plenty of information without possessing material a model can use reliably.

The customer's problem also extends beyond the first training run. A useful system needs evaluation and revision. At GEOINT 2025, product leader Tim Klawa emphasized aligning training-data attributes with the needs of operators and warfighters. The interesting question is how well the examples represent the situation in which the machine will eventually be used.

02 / Silicon Valley, with a federal address

The name carries some history. Figure Eight, previously CrowdFlower, began in San Francisco in 2007. Appen announced its acquisition in March 2019, offering $175 million upfront and as much as $125 million more based on performance. That was a purchase of Figure Eight, rather than a funding round for the federal subsidiary.

Figure Eight Federal LLC has a separate date: August 2019 in federal registration data. By April 2020, Appen was publicly describing a government-focused operation supporting defense, intelligence, disaster recovery and law enforcement. It also described an Esri partnership aimed at humanitarian work during natural disasters. The applications stretched beyond identifying things in military imagery.

The organizational distinction matters. F8F is described in partnership announcements as Appen’s FOCI-mitigated arm, referring to foreign ownership, control or influence. For a U.S. government buyer, who owns a supplier and how sensitive work is handled belong in the same conversation as product features. Appen’s 2025 annual report still lists the federal LLC as wholly owned.

03 / Three partners, three different problems

F8F’s partnerships offer a useful map of the business. Latent AI addresses model optimization and deployment at the edge. The collaboration announced in January 2024 paired F8F’s labeling and workflow capabilities with Latent AI’s machine-learning operations technology. Preparing examples and making a model practical to deploy are connected jobs, requiring different expertise.

Enveil addresses another obstacle: querying sensitive interests against outside data. Its October 2024 integration connected encrypted search with F8F’s HYDRA AI Analytics Studio and Global Location Data Feed. The stated aim was to let analysts search across data holdings while protecting their queries, avoiding the need to move entire collections into each customer environment.

HYDRA broadens the picture beyond labeling. Its portfolio concerns geospatial analytics and patterns in human movement, supplying material for situational awareness and decision support. That makes the Enveil pairing intelligible: possessing useful location data and being able to search it privately are distinct requirements.

Microsoft addresses the surrounding infrastructure. In May 2025, F8F announced an Artemis integration with Azure services for governance, data pipelines, MLOps and analytics. The announcement described modular components and consumption-based purchasing. For buyers, the appeal is understandable: improve a pipeline without committing every part of it to one permanent supplier.

04 / The number that needs a footnote

Federal technology announcements often come with large dollar figures. F8F’s February 2022 release announced selection in a partnership for the Joint Artificial Intelligence Center’s $249 million blanket purchase agreement. It provided a route to supply AI capabilities. The figure describes the procurement vehicle, not a $249 million payment to F8F.

There is a more revealing financial question underneath: what should the next increment of training data cost, and what improvement should it buy? F8F, the Training Data Project and Technomics explored that relationship in a paper linking acquisition strategy, risk and AI outcomes. ICEAA named it the 2025 category winner for Management, EVM, Software and Agile.

The paper’s premise is that programs need measurable criteria that can follow work across contracts, from training data through testing and operations. Its warning concerns acquisition frameworks that lag behind rapidly changing capabilities. This puts the discussion in the territory of budgets, schedules and decisions about where another dollar will help.

05 / Borrow the discipline

“Can you trust AI with the decision?”Tim Klawa / KnowledgeNile, 2025

F8F operates in a competitive market. Scale’s Public Sector Data Engine also offers annotation, curation, evaluation and human expertise. F8F’s particular proposition combines specialized sensor formats, federal operating arrangements and integrations around the labeling workflow. Those are features for buyers to examine against their own requirements.

The practical lesson to borrow is to define the decision before ordering the labels. Identify the difficult cases, involve people who understand them, then connect quality measures to performance and cost. The approach depends on representative data and suitable experts. It cannot demonstrate performance in conditions the evaluation never covers. F8F’s business begins with that small, stubborn act of definition: deciding what a machine actually needs to learn.