Clinical Workflow Assistant

The initial architecture was AWS Lambda — stateless, cheap, appropriate for the summarisation phase. Chat changed the equation. WebSockets require persistent, stateful connections. Lambda does not support this reliably. The move to Fargate solved statefulness but introduced disproportionate cost at our scale. A single EC2 instance gave equivalent capability with more control and lower overhead. Two infrastructure migrations, both driven by what the product required at each stage.

The first chat implementation was built entirely on the raw GenAI SDK — including a custom tool registry and tool-calling infrastructure written from scratch. It worked, but as clinical flows became more complex, the maintenance burden compounded. Retry logic, state management, and error handling all had to be manually coordinated across an expanding codebase.

LangGraph provided graph-based flow management with built-in retries, checkpointing, and idempotency. The migration replaced the custom infrastructure with framework-level primitives. What followed — multi-step encounter generation pipelines, conditional flow branching, two-pass AI orchestration — would not have been maintainable in the original approach.

Most ambient scribe tools are web-based. A mobile-native encounter module was an underserved gap in the clinical workflow space. The client needed something that fit how clinicians actually work — interrupted, context-switching, time-pressured.

My first framing was a sequential wizard: select patient → record → generate notes. Logically coherent, wrong for clinical reality. The reframe: everything optional, everything additive.

The final structure has three screens. An encounter screen where patient selection, document attachment, and recording are all independent — a doctor in a rush can skip straight to recording. A recording screen where audio can be recorded live or uploaded, documents can be attached as context, and generation can happen from documents alone. A review screen where additional inputs can be added and notes regenerated across multiple templates — SOAP, DAP, H&P, and others — as many times as needed.

The client approved it without revision. Automatic patient detection and document attachment support were added in the two weeks following initial delivery.

The patient memory system needed to handle corrections — when a clinician updates information that contradicts what was previously recorded. Every prompt-level technique failed identically: flat prohibitions, XML sectioning, synonym banlists, scratchpad reasoning, few-shot examples, and pipeline splitting. The model kept treating corrections as clinical timeline events rather than resolving them.

The fix was to stop fighting the model weights. Correction logic moved entirely into deterministic application-layer code via LangGraph node interception — corrections are stripped before the model ever sees them, resolved directly in the application, and applied as targeted field patches. The principle generalises: prompt engineering has hard limits for deeply trained behaviors. When every surface-level technique fails identically, the fix is architectural, not instructional.

Previously, patient-scoped chat only carried demographic context. The model knew who the patient was, not what their case was.

Patient memory restructured this by maintaining a clinical context object per patient, synthesised entirely from encounter documents, audio transcripts, and facts provided directly in the chat. The model now enters each session with grounded clinical knowledge, which also populates the patient profile to give clinicians an instant clinical snapshot.

A comprehensive UX audit aligned UI friction and contextual clarity directly with the stakes of the operation.

Destructive actions and errors now demand explicit modal confirmations or blocking acknowledgements, while low-stakes updates use non-intrusive toasts. Similarly, high-stakes operations feature immediate inline hint text, whereas low-stakes operations rely on unobtrusive info icons, ensuring critical information is impossible to ignore without cluttering the core workflow.