CLINICALBRIEF'Sign in
Roadmap

Roadmap

Where ClinicalBrief stands, phase by phase. Statuses are honest: “prototype” means it works but is deliberately simple. Planned phases carry no dates.

Progress

Phases at a glance

  • 10Done
  • 1Next
  • 4Planned

Every phase is independently code-reviewed before it counts as done.

Timeline

Phase 0 to Phase 13

  1. Phase 0done

    Audit

    A full review of the starting codebase: what worked, what was claimed but not real, and what had to be rebuilt before anything else.

  2. Phase 1done

    Data

    Synthea FHIR R4 import into one canonical model (patients, encounters, notes, diagnoses, medications, allergies, procedures, observations). Dry run, per-record validation, rejected-record log and idempotent re-import; every record keeps its source and provenance.

  3. Phase 2doneprototype

    AI note pipeline

    Rule-based extraction: dictionary matching with negation and family-history handling, plus an extractive summary copied from the note. There is no calibrated confidence, so confidence is shown as empty rather than as a number.

  4. Phase 3doneprototype

    Knowledge layer

    A patient-centred graph of conditions, findings, medications and allergies. AI findings appear only after human review, and links between them are labelled as co-occurrence, not clinical fact.

  5. Phase 4doneprototype

    Grounded Copilot

    Questions answered by a local model grounded in one patient's record. Patient data never leaves the machine; if the model is unavailable, labelled rule-based answers are used instead.

  6. Phase 5done

    FHIR R4 export with validation

    A transaction bundle built from the canonical record, checked on every export for R4 types, cardinality and resolving references. Structural R4 only: not validated against Australian profiles or terminology.

  7. Phase 6done

    Human review and governance

    A review queue where people approve, reject or edit each AI finding against its evidence sentence, with every decision audited. A governance page shows pending work and decisions; nothing AI-extracted is used as fact until a person has reviewed it.

  8. Phase 7done

    Clinical workspace

    The patient record, notes and layouts that tie the pieces together in one place to work from.

  9. Phase 8adone

    Record search

    Postgres full-text search over every note and the structured record, with patient, type and date filters. Keyword-based, not semantic, and scoped to the patients a user may access.

  10. Phase 8bplanned

    Semantic search

    Embedding-based search, kept only if it beats full-text search on a fixed set of questions.

  11. Phase 9done

    Observability

    A System health page for admins and auditors: live checks of the database, local model and search index, plus request timing, error rates and request ids. Kept in memory per backend process, so history resets on restart.

  12. Phase 10next

    Security hardening

    A focused pass over authentication, access rules and data handling to close remaining gaps.

  13. Phase 11planned

    Demo

    A guided way to see the product working end to end on synthetic data.

  14. Phase 12planned

    Industry readiness

    Documentation a technical reviewer would expect: how it is built, how to run it, and what its limits are.

  15. Phase 13planned

    Portfolio

    Presenting the project and what was learned building it.

Not built yet

Deliberately missing

  • Accuracy evaluation on a labelled set. No accuracy figure is shown anywhere until it is measured.
  • Semantic search (Phase 8b).
  • A de-identified research view.