Medical Context Engineer
Build reliable HealthOS medical records from documents and wearables, with source links, provenance, review, and chat retrieval.
RemoteUSD 252,500/year
Anistratenco is building a longitudinal health platform. Health Assistant brings conversations, records, and goals into a context people can inspect and correct. The underlying systems must support more products as that context grows. This full-time role is remote within the United States.
The mandate
Build document and wearable imports into one typed medical record store; keep explicit personal notes separate.
Own deterministic source validation, ambiguity review, bounded record retrieval, context assembly, and evaluation. Preserve provenance, observation time, conflicts, correction, and deletion boundaries.
Show us the work
Send two files: a complete project case study and a proposal for this role. No CV or cover letter. Use public evidence and synthetic examples where needed. AI tools are welcome; the technical decisions, evidence, and proposed execution must withstand review.
Project case study
Document one substantial project from the original problem through launch and subsequent results. Explain the team, your own scope, timeline, constraints, architecture or operating stack, key decisions, implementation, and what changed along the way. Include work you can substantiate, a dated baseline, the measurement method and denominator, and the measured result. Discuss a failed approach or tradeoff. Label estimates and evidence you cannot verify.
Show a complete context, retrieval, or memory system you helped take from architecture to production operation. Detail ingestion, schema, model and open-source components, storage, orchestration, retrieval, evaluation, deployment, and incident handling. Explain your contribution, design revisions, failure cases, and results for quality, latency, cost, and stale or contradictory records.
Role proposal
Infer the current HealthOS and context path from source; cite evidence and mark unknowns. Design source-bound imports from PDFs, images, and wearable observations into typed records, with deterministic quote checks, a review queue for ambiguity/conflicts, bounded record tools for chat, and a small turn-start summary. Include a simple diagram, storage/index boundaries, deployment topology, operational runbook, observability, and quality/latency/cost/isolation evaluations. Keep model suggestions separate from accepted medical facts.