Problem
Healthcare teams need faster access to knowledge and assisted workflows, but unconstrained LLMs invent facts, ignore policy, and fail audits. The challenge was to design GenAI that is grounded, reviewable, and operable in a regulated setting.
My role
- Owned GenAI / agentic workflow design for selected product use cases.
- Partnered with product and domain stakeholders on evaluation criteria and human-in-the-loop checkpoints.
- Shaped retrieval, tool-use, and safety patterns so demos could become operable features.
Architecture (shape)
- Retrieval first — answers and actions tied to approved sources, not open-ended generation.
- Agent / tool layer — constrained tools for workflow steps; explicit handoffs for review.
- Guardrails — policy-aware prompts, grounding checks, and escalation paths for low-confidence outputs.
- Observability — traces for prompts, retrieval hits, and failures so operators can debug production behavior.
Approach
- Framed use cases around retrieval and tool use, not open-ended chat.
- Designed RAG and agentic patterns with explicit sources and human-in-the-loop checkpoints.
- Iterated evaluation against real workflows, not generic LLM benchmarks alone.
Trade-offs
- Preferred narrower, auditable assistants over broad chat — lower “wow,” higher trust.
- Accepted extra latency from retrieval / checks when it improved grounding.
- Kept humans in the loop for high-impact actions instead of full autonomy.
Evaluation
- Groundedness / citation fidelity against source corpora.
- Workflow success with domain reviewers (not only offline NLP scores).
- Failure modes: refusal quality, escalation, and safe degradation.
Outcome
Delivered production-oriented GenAI capabilities that improved knowledge access and assisted operations while keeping safety and grounding first-class.
- Metric slot: e.g. grounded answer rate / human override rate — fill when shareable
- Metric slot: e.g. time-to-answer for a known workflow — fill when shareable
- NDA: Exact client metrics available in conversation under NDA.