Cognitive Systems: Receipt OCR, RAG Knowledgebase, Assist

The client is the leading HSA and health-finance custodial administration platform with millions of HSAs and total accounts, and tens of billions in HSA assets. The business generates over a billion dollars in annual revenue at strong EBITDA margins. Taller's engagement at this client is one of the broadest multi-pod engineering footprints in the portfolio outside the global payments company organization. The three cases below describe what Taller has built or contributed to across the multi-tenant cloud modernization, the integration substrate for vendor-provided trust capabilities (Plaid, passkeys), and the team's same-client AI tooling work.

The HSA platform company's cognitive-systems portfolio — a receipt OCR claims product, a RAG knowledgebase over hundreds of curated resources, a personalization engine, the Assist portfolio for analytics, agentic AI through their agentic AI provider — targets the decision moments where members engage with the platform. Production AI in a regulated context requires evaluation gates, audit-traceable retrieval, and release pipelines that mirror production without affecting live adjudication.

Taller's same-client AI engagement at this client is limited to GenAI Copilot tooling used inside the Identity team to generate C# and Node.js code for identity services. The broader production-AI capability Taller carries is anchored elsewhere: the major US airline portfolio of generative-AI engagements — productivity AI with hybrid search, multi-query retrieval, and RAGAS evaluation; RAG for aircraft manual version comparison with LangChain, LangGraph, AWS Bedrock, knowledge graphs, and re-ranking; an enterprise generative-AI framework supporting multi-agent workflows and graph-based RAG — serving as the capability reference for what production AI looks like in a regulated, document-heavy operating context.

Same-client AI engagement at this client covers GenAI Copilot tooling adopted inside the Identity team for SDLC acceleration. No same-client delivery on the receipt-OCR claims product, the RAG knowledgebase, or the personalization engine products themselves.

The HSA platform company's cognitive-systems outcomes — approximately seventy percent average time savings in claims entry; more than fifty percent of claims processed in under two minutes; an eighteen percent satisfaction lift in the 2024 release to over a million members; 368,000 members taking personalization actions through the personalization engine with a ninety percent higher likelihood of increasing contribution amount and a ten percent higher likelihood of beginning contributions; one early-adopter state reporting significant HSA enrollment growth in twelve months — are the HSA platform company's product outcomes earned by the HSA platform company's product portfolio and their agentic AI provider's agentic AI. Capability fit between Taller's regulated-AI work at this airline and the HSA platform company's cognitive-systems strategy is real; same-client deployment on the cognitive-systems products themselves is not in Taller's documented scope.