Loading...
Client: American Legion Ambulance — Amador & Calaveras Counties, California
Engagement: AI/ML solution design, build, and production operation on AWS
American Legion Ambulance has run the 911 contract for two rural Sierra-foothill counties since 1929. Its crews work two-paramedic rigs on long-transport routes where the receiving hospital can be 45 minutes away and the right clinical decision has to come from the crew, not a phone call. Every one of those decisions traces back to a ~300-page county Field Treatment Protocols binder — dosing tables, pediatric charts, and procedural checklists that have to be exactly right and instantly findable.
A static PDF on a phone doesn’t meet that bar. Dosing tables are images to a text search. Protocol titles don’t follow one convention. Single protocols split across page breaks. The crews trusted the binder precisely because they didn’t trust the search.
Jacobian Engineering designed, built, and now operates a production retrieval-augmented generation (RAG) assistant for the rig. A paramedic asks, in plain language, for a weight-based pediatric epinephrine dose or a stroke-alert checklist — and gets a synthesized, citation-backed answer in seconds, with a tap-through to the exact source page.
The system is Bedrock-native end to end. Amazon Bedrock Knowledge Bases manages retrieval over an Aurora PostgreSQL pgvector store; Amazon Nova Pro generates, Cohere Rerank sharpens retrieval, and Titan Embeddings v2 vectorizes. The hard engineering lives in ingestion: a layout-aware Docling pipeline preserves clinical tables intact and uses a vision-language model to recover section titles where the binder’s own formatting breaks down — the difference between a RAG demo and a tool a paramedic will trust at 2 a.m.
Around that core, Jacobian shipped an enterprise security and observability posture on day one — GuardDuty, Inspector, Security Hub, Macie, Config, CloudTrail, KMS — plus a global kill switch, full conversation audit, and paged alarms before the first crew ever saw the interface.
The binder now updates at the speed of a shift change: a medical-director revision lands in S3 at noon and is answerable, with a fresh citation, by the next crew on. New paramedics ramp faster; veterans use it as a cross-check. Adoption isn’t driven by the model being clever — it’s driven by every answer being checkable against the original page.
It’s a working sample of how Jacobian brings enterprise-grade AI/ML engineering to an SMB-scale, mission-critical workload — and operates it in production.
Read the full engineering story → The complete American Legion Ambulance case study on TrustEdge
Challenge: Put a 300-page county EMS protocol binder in every paramedic’s hand with citation-backed, instantly findable answers a crew can trust at 2 a.m. Solution: A production, Bedrock-native RAG assistant — Nova Pro, Cohere Rerank, Titan Embeddings v2 over pgvector — with a layout-aware ingestion pipeline that preserves clinical tables and recovers binder section titles.