AI Engineer
Mentriix · Oct 2023 – May 2026
- Member of the client's AI Center of Excellence, building enterprise GenAI products on Amazon Bedrock for utilization-management nurses, care managers, and clinical reviewers using Python (LangChain, LangGraph, FastAPI), Node.js / TypeScript (NestJS), and React / Next.js, deployed on AWS Lambda, Step Functions, and ECS Fargate via AWS CDK.
- Built core components of a Clinical Document Intelligence platform using a RAG pipeline with Amazon Textract, Claude 3.5 Sonnet on Bedrock, and Bedrock Knowledge Bases on OpenSearch; processes medical records, prior-authorization attachments, claims documentation, and provider notes into structured clinical summaries, significantly reducing clinical review time.
- Engineered the document-ingestion and indexing pipeline (S3 → Step Functions → Amazon Textract → semantic chunking → embeddings) with incremental re-indexing of Bedrock Knowledge Bases as new medical records, prior-authorization attachments, and medical-policy documents arrived.
- Developed a Utilization Management Copilot on a LangGraph multi-agent architecture (retrieval, evidence-grading, medical-policy-adherence, and citation agents) with hybrid OpenSearch and pgvector retrieval across medical policies and member clinical history, improving medical-necessity answer accuracy on internal evaluation sets.
- Contributed to an Agentic Prior-Authorization Assistant on Amazon Bedrock using Model Context Protocol (MCP) servers that expose tools for medical-policy lookup, member clinical history, and benefit / coverage checks; the agent plans multi-step medical-necessity reviews and drafts prior-authorization determinations with cited rationale and reasoning traces.
- Designed, versioned, and A/B-tested prompts per agent role
- system and few-shot templates
- enforcing structured JSON outputs through Bedrock tool-use / function calling for reliable downstream parsing and orchestration.
- Tuned retrieval quality using Amazon Titan embeddings, metadata-based filtering, and a reranking stage over hybrid OpenSearch and pgvector search, improving context relevance and reducing irrelevant retrievals in evaluation runs.
- Implemented the GenAI evaluation and guardrails layer with self-hosted Langfuse for LLM tracing, a Ragas and DeepEval harness integrated into CI, and Amazon Bedrock Guardrails for PHI redaction, jailbreak filtering, and toxicity checks, helping reduce the hallucination rate on production traces.
- Curated golden and regression evaluation datasets with clinical SMEs (utilization-review nurses and medical directors) and automated nightly evaluation runs in CI to catch accuracy and faithfulness regressions before release.
- Optimized inference cost and latency through model routing (lightweight models for extraction and classification, larger models for reasoning), prompt caching, and parallel tool execution, with token-usage dashboards for per-product cost tracking.
- Implemented data-governance and compliance controls for a regulated healthcare environment
- HIPAA-aligned PHI handling, KMS encryption, VPC-isolated Bedrock endpoints, and least-privilege IAM
- and authored model cards and documentation supporting internal Model Risk Management (MRM) and clinical governance reviews.
- Developed backend services for the copilot suite in Node.js / TypeScript (NestJS) with streaming Server-Sent Events (SSE), Cognito-based OAuth2 / OIDC single sign-on, role-based access control, and CloudWatch audit logging aligned with internal Model Risk Management controls.
- Built front-end features in React and Next.js (TypeScript, Redux Toolkit, TanStack Query, Tailwind CSS) for clinical-reviewer and care-management users, including a streaming chat interface with inline citations, document preview, and user-feedback capture wired into the evaluation loop.
- Deployed services via AWS CDK (Lambda, Step Functions, API Gateway, S3, DynamoDB) with ECS Fargate for long-running agent workloads; configured GitHub Actions CI/CD pipelines with Datadog and OpenTelemetry observability, improving deployment cycle times.
- Collaborated in Agile teams with product managers, designers, and business stakeholders through sprint planning, daily standups, code reviews, and demos to deliver features iteratively.
