Four state programs,and one codebase to serve them.

Before this, a new state program meant a copy of the repository: its own prompts hard-coded, its own guardrails, its own database bindings, and its own defects from the moment the copy was taken. A MongoDB-backed admin control plane replaced that. A program is a record — the prompts it uses, the rails it is held to, the databases it is permitted to read — and all seven services resolve that record per request. Onboarding the next program became filling the record in: from multi-week custom development to under two days of configuration.

A caseworker's question is not one kind of question. What the policy says about a placement move is retrieval; how many cases opened last month is SQL; find the note that mentions the school transfer is a search index; what can you actually see is the system describing itself. A ReAct agent classifies every inbound query and dispatches it across all four through tool calling on AWS Bedrock, inside a bounded ten-iteration loop so that no question can run away with the budget. In front of it sits a lightweight pre-router that resolves greetings, closings and follow-up intent before the agent is invoked at all — about 30% of queries never open the loop, and never cost what the loop costs.

Owning the platform's AI spend meant separating the work that needs reasoning from the work that only needs structure. Structured query generation is high volume and narrow, so it runs on self-hosted Hugging Face TGI; reasoning and summarisation stay on managed Bedrock, where the quality is worth the rate. Embedding generation came off the request path entirely and onto CPU inside the batch worker, which keeps GPU capacity out of ingestion. Together the routing took inference spend down about 40%, with no change a caseworker could see.

Documents arrive as PDF, DOCX and XLSX and leave as vectors. The upload is asynchronous from the first hop — FastAPI writes to Amazon S3 through aioboto3, a message goes to SQS, and a containerised worker picks it up, parses with docling and PyMuPDF, and writes to Milvus. Every message carries its state in PostgreSQL and every retry is idempotent, so a document that fails is a row somebody can explain rather than a file that quietly did not arrive. Caseworker document search time fell about 70%.

Public-sector records do not get to leak between programs, and the design starts there rather than adding it later. Retrieval runs over one unified 1024-dimension Milvus collection with program-scoped filtering, so a query is bounded before it is scored, and CrossEncoder reranking takes it the rest of the way — retrieval precision up about 35%, unsupported responses down about 45%. Around it sits the record a review asks for: conversation memory in Redis, and queryable audit logs of the model's reasoning and its tool selection.

A platform assembled service by service has a shape nobody has written down. Cross-repository technical due diligence produced it: a system contract map and a dependency graph across all seven services, and with them the things a map makes visible — two production integrations that were broken, five code paths nothing could reach, and secret-management exposures in every repository. It ended as a prioritised remediation roadmap put in front of engineering leadership, ordered by what breaks first rather than by what is easiest.