Hathor
Hathor is an autonomous clinical reasoning agent for cross-border vaccination reconciliation.
Core trick
Vision confidence gates route low-confidence fields to clinician review, and a deterministic WHO-DAK validator vetoes unsupported clinical recommendations.
Description
Hathor is an autonomous clinical reasoning agent for cross-border vaccination reconciliation. A child arrives in Egypt with a Nigerian (or Sudanese, Syrian, Yemeni…) vaccination card; Hathor reads the card, resolves trade-name combinations to canonical antigens, checks every dose against the destination schedule, and emits a visit-by-visit catch-up plan — all with the agent deciding its own tool path, no hardcoded pipeline. Built on the Claude Agent SDK with eight in-process MCP tools and two safety loops: a per-field vision gate (anything < 0.85 confidence routes to clinician HITL review) and a per-recommendation rules-engine gate (a deterministic Python WHO-DAK validator vetoes any clinical claim before it reaches the UI or FHIR bundle). The clinician retains final authority and every override is logged to a FHIR Provenance resource. No open-source tool exists for autonomous cross-schedule reconciliation today; the closest prior work (AI-VaxGuide, arXiv 2507.03493) is single-country Q&A.
README excerpt
# Hathor An autonomous clinical reasoning agent that reconciles a child's vaccination history across national immunization schedules. --- ## Why Millions of African families move between countries every year — within the continent, for work, study, or safety. A child vaccinated correctly under one national schedule often arrives in a new country where no one can quickly tell the family which doses count, which are missing, and what is needed before the child can enrol in school or nursery. The schedules look similar on paper but diverge in important ways: the antigens covered by combined products, the ages of administration, the vaccines that are routine in one country and not another, and the strict age windows on products like rotavirus. I am an Egyptian physician. I built Hathor because I kept running into this problem myself and in my clinical work — and because no open-source tool exists for autonomous cross-schedule reconciliation. The closest prior work (AI-VaxGuide, arXiv 2507.03493) does clinician Q&A over a single country's guidelines, not reconciliation across schedules. --- ## How it works Hathor is built on the Claude Agent SDK with extended thinking enabled. Eight custom clinical tools are exposed via an in-process MCP server: card extraction, age computation, vaccine equivalence lookup, interval validation, per-dose validation, schedule retrieval, gap analysis, and catch-up scheduling. The agent decides which tools to call, in what order, based on what