Production Clinical AI Platform
Heavy contributor to a production clinical AI system — Ruby backend, AWS Bedrock LLM pipeline, Zoom transcript processing, shipped and supported live.
Overview
This is a de-identified production healthcare AI system. Joined an established architecture mid-build and contributed heavily across the full stack — not a greenfield project, but the harder challenge: coming into a live codebase, understanding it quickly, and shipping reliably at production quality. The platform converts clinical conversations into structured charting notes. The core pipeline: a Sidekiq background job fetches a Zoom VTT transcript from S3, maps the appointment's configured note template to a set of custom module fields, renders a structured ERB prompt, sends it through an internal AWS Bedrock proxy (with up to three retries), parses the LLM's structured response, normalizes field-type-specific values (including checkbox option matching), and persists the generated form answers. Every step emits Datadog metrics and updates an Observatory record that the internal support dashboard reads for troubleshooting. Work covered the complete delivery arc: implementing new pipeline features, writing and maintaining RSpec coverage, debugging production regressions, owning release coordination, and providing live-site support. The job was not just to write code but to make the system shippable — tracking down intermittent failures in the transcript fetch path, hardening the LLM retry loop, and being accountable for what went out the door. This is the kind of engineering that does not appear in demos: owning quality and reliability on a production AI system in a regulated domain, under deadline, as a contributor to an existing team. *Code examples on this page are representative illustrations of the architectural patterns used — they are not actual proprietary source code from the production system.*
Technical Stack
AI / LLM
- ▸AWS Bedrock
- ▸ERB prompt templates
- ▸Structured output parsing
- ▸Custom module mapping
- ▸Retry with backoff
Backend
- ▸Ruby on Rails
- ▸Sidekiq
- ▸ApplicationService pattern
- ▸Dependency injection
- ▸RSpec
Infrastructure
- ▸AWS S3
- ▸Datadog metrics
- ▸Mixpanel events
- ▸CircleCI
Key Features
Sidekiq background job pipeline triggered automatically after a Zoom call ends
Zoom VTT transcript fetching from S3 with stub transcript support for local development
Dynamic prompt construction via ERB templates: system prompt + user prompt rendered against appointment context
Custom module map builder mapping each charting note field to its LLM prompt section by appointment type
AWS Bedrock LLM invocation via internal proxy with up to three retries on transient failures
Structured answer parsing with checkbox option normalization (fuzzy-matching model output to valid option values)
Per-task Observatory telemetry tracking provider, patient, status, and failure reason for every generation attempt
Datadog metrics on every pipeline stage plus Mixpanel event on successful note generation
Code Examples
Technical Challenges
Ramping into a live production codebase quickly enough to ship features on the team's existing release cadence
Achieving reliable test coverage for a pipeline whose core outputs are LLM-generated and non-deterministic
Debugging intermittent transcript fetch failures where root cause was upstream S3 upload timing
Hardening the LLM retry loop to handle Bedrock throttling without surfacing errors to end users
Maintaining production reliability in a healthcare context where charting failures have direct clinical consequences