Context
A government energy organization was building an internal knowledge assistant over its own document estate, so staff could get answers out of material that previously had to be found and read. The document estate was not tidy: mixed formats, and a long tail of scanned material that carried real content and survived naive extraction badly. When we came in, the project was behind schedule.
Constraints
A compliance environment covering controlled unclassified information, which put security review on the critical path rather than at the end of it — every ingestion path, storage decision, and inference location needed auditing and approval. The architecture had to fit an Azure and .NET platform the organization already ran and had to keep running, rather than arriving as something adjacent to it. And the retrieval had to hold up against the real corpus rather than a curated sample.
RegainFlow's role
RegainFlow provided technical leadership and assessed business value for the organization's IT. We led the architecture for document ingestion and Elastic-based retrieval, stood up on-premise inference and the operational practice around it, and worked inside the organization's stack, security process, and delivery cadence. We did not own the programme; we were brought in to help it land.
What RegainFlow engineered
- 01
Assess
Establish where the value actually was for this organization and which document sets would carry it, then set the technical direction against a schedule already in deficit.
- 02
Ingest
Build an ingestion path across a mixed estate, including scanned material, that turns the organization's own documents into a form retrieval can use.
- 03
Secure
Work alongside security on auditing and approval so the compliance review ran with the build instead of behind it, covering ingestion, storage, and where inference is allowed to happen.
- 04
Operate
Stand up on-premise inference with the deployment practice and usage measurement around it, so the team that runs everything else can run this too.
Selected artifacts
Reconstructed, not screenshotted.
Where a document goes
- Mixed estate
- Ingestion
- Retrieval index
- On-premise inference
- Answer
What the client received
- An ingestion path across a mixed document estate, scanned material included
- An Elastic-based retrieval architecture sized to the corpus
- On-premise inference running inside the compliance boundary
- Deployment practice and usage measurement for the platform team
- A business-value assessment tying the build to what the organization needed
Outcome
An internal knowledge assistant in the organization's own environment, answering questions from its document estate on infrastructure its team already knows how to run. The project came back on track, secured further funding, and found an unusually eager set of participants — adoption ran ahead of what the programme had planned for.
What changed next
The system is evolving from single-turn question answering toward multi-step agentic workflows — moving from answering a question to carrying out the work the answer implies.