About the role
Our stack: Python for the AI services, Amazon Bedrock (Mumbai region) for large language models (LLMs), TypeScript and NestJS for the core, PostgreSQL 16, Amazon Web Services (AWS) — ECS Fargate, S3, EventBridge, OpenSearch.
What you will build • The plug-in runtime and registry: manifest and scopes, event subscriptions, and a per-hospital install and kill switch. • The LLM gateway: model routing, prompt and version management, cost and token controls, redaction of identifiers before anything leaves the boundary, and full auditing of every request. • The three launch plug-ins: 1 Discharge Summary Assistant — drafts a summary from the encounter record for the doctor to edit and sign. 2 Clinical Scribe — turns a consultation conversation into structured notes. 3 Coding Assist — suggests International Classification of Diseases (ICD) and procedure codes from the note. • The evaluation harness: a golden dataset, regression runs on every prompt change, and dashboards for acceptance rate and edit distance per hospital.
Responsibilities • Treat every output as a suggestion. Nothing your layer produces is ever written directly into a clinical or financial record without a human accepting it. • Own quality measurement. "It looked good in testing" is not a release criterion here — acceptance rate is. • Work with the clinical advisors to sign off outputs before a plug-in is enabled for a hospital. • Keep patient data handling defensible: what is sent, where it is processed, what is retained, and the audit record for every call. • Control cost per plug-in per hospital, and report it monthly.Requirements • Python, and comfort building and operating a production service — not only notebooks. • Practical LLM application patterns: prompt design, structured and schema-constrained output, retrieval-augmented generation (RAG), function calling, streaming, and evaluation. • Has shipped at least one LLM feature to real users, and can describe how quality was measured and what was changed as a result. • Amazon Bedrock, the Anthropic API or the OpenAI API in production. • Pragmatism: you choose the smaller model and the simpler prompt when it does the job.Good to have • Speech-to-text for Indian-accented English and code-switched speech — the Clinical Scribe depends on it. • Medical coding, clinical text, or de-identification work. • TypeScript, so you can meet the core team halfway on the gateway. • Evaluation tooling and experiment tracking.