Scan, Extract, Translate: how A4A turns paper into coded history

IMDS and REMIS hold the maintenance history of the fleet. The qualitative detail that would shorten a troubleshooting cycle sits in free-text narrative fields typed under time pressure, full of typos, non-standard acronyms and shop shorthand. Analytical tools cannot parse it.

The Government's question about a thirty-day delivery is not whether Edgent can write software quickly. It is schedule risk: what happens to the milestone dates when the hard parts turn out to be hard.

A4A answers that by not starting at zero. The orchestration engine, the governed storage, the retrieval substrate and the authorization model are already built and already running. That head start is the schedule. The thirty days is spent on aviation, on the corpus, the ATA taxonomy and the tuning, instead of on the infrastructure that would otherwise consume most of it.

The components, and what each one actually contributes

Each is marked shipped or design intent on the same basis as the pipeline table above, and where a component is real but is not yet doing the A4A job, that is said rather than blurred.

The platform

Every Edgent product is built the same way: a domain model with ports and adapters, so storage, messaging and inference are interfaces the domain is written against rather than dependencies wired through it. Two consequences matter to a compressed schedule. Multi-tenancy is structural rather than a filter someone has to remember to apply: the tenant is bound to the credential, not to a header a caller can set, and it becomes part of the storage path itself. Authorization happens at a gateway, before a request reaches any application code, where a Keycloak-issued token is validated and a relationship-based access decision is made in OpenFGA. A4A therefore ships with no tenancy or authorization logic of its own, and that is the part of a new federal system least forgiving of a short timeline.

Shipped

Stated precisely: the policy engine intended to sit alongside OpenFGA is written and unit-tested but is not yet in the request path. It is not claimed as enforcing.

VForce Flow

The orchestration engine. Pipelines are drawn in a visual designer and kept as YAML, validated by flowc and applied and published to the running engine by flowctl in CI, so changing a pipeline is a reviewed merge rather than a console edit. Durable execution runs on Temporal in the deployed configuration, behind an engine port, so retries, timeouts, conditional branches and run history belong to the platform rather than being rebuilt per product. The branch that turns a silently empty OCR result into a failed run, described under pipeline, is a few lines of that YAML rather than a code change.

Shipped

pixelMapper

Visual similarity search. A self-hosted vision-language model embeds an image into the same vector space as text, so a photograph can be the query and a nearest-neighbor search returns what it resembles rather than what it is described as. For A4A this is the route to identifying a part or a form from a photograph instead of from a typed description.

Shipped

In service today against a different corpus, on Edgent's forensic image-matching path. Nothing on this page uses it. The demonstration above reads a page with vision OCR and then retrieves over text; image-as-query is a capability Edgent operates, not one exercised here.

The lakehouse

Governed storage on Apache Iceberg, with the bronze, silver and gold layer a first-class property of a table rather than a naming convention. Each tenant gets its own namespace root, so a tenant boundary is a fact about the storage path and not only a predicate someone appended to a query. The audit log is hash-chained: every entry carries the hash of the entry before it, so an edit or a deletion breaks the chain and becomes detectable rather than merely prohibited. Every page A4A reads lands as a governed row, and the record of who read what is kept under the same controls, in the same place. The audit page publishes the live chain proof for this demo: whether the chain is intact, right now, on demand.

Shipped

Vector similarity search

Retrieval over prior maintenance records is what turns a new write-up into a list of comparable historical cases and what fixed them. It runs on a self-hosted BAAI/bge-m3 embedding service on Edgent's own GPU, open weights, deliberately local so the vector space never leaves the deployment. It is the same service that already backs retrieval in the platform: the figures under measured results were produced by calling it, not by a model stood up for this demonstration.

Shipped

Stated precisely: the index for the document path is held in memory today and rebuilt from the gold table after a restart. A durable index is already wired for other collections, so this is configuration rather than new work, but it is not what the document path uses at present.

Where the thirty days goes

Read down the last column: the infrastructure rows are settled and the open rows are aviation. That is the difference between a thirty-day estimate and a thirty-day commitment. Configuration and content on a running system can be scheduled; standing up orchestration, governed storage and a retrieval substrate inside thirty days could not be.