About the project
Prognosia - writes the visit for you, and lets no error through
Doctors spend more time on paperwork than seeing patients (~49% of the day in the EHR vs 27% face to face), and 30% of malpractice suits involve a communication failure. Cloud AI scribes exist, but no AI enters a hospital without passing compliance.
Prognosia is designed in reverse: requirement first, feature second. Everything runs on the clinician's device. Data never travels.
What it does
- The doctor picks a patient (or registers one in seconds), presses one button and talks to the patient as usual, or pastes/types what was said.
- Prognosia transcribes the visit, drafts a structured SOAP note, and proposes what to write into the EHR.
- Before anything can be signed, a deterministic safety engine cross-checks every order against the chart (history, allergies, current meds). If the plan contradicts the record, the order is withheld with the evidence on screen: chart says severe asthma, visit says "start propranolol", signature blocked until the plan is fixed.
- Nothing enters the medical record without the clinician's signature.
How it works
- STT on-device: Whisper (WHISPER_BASE_Q8_0 + Silero VAD) via the QVAC SDK, with a post-STT clinical lexicon for drug-name robustness.
- Note extraction on-device: Qwen3-4B (Q4_K_M) generates the SOAP note as Pydantic-validated JSON, with retry-with-feedback and explicit refusal. It never invents medication.
- Safety decision without an LLM: deterministic rules evaluate the raw transcript as well as the extracted note, so a near-miss is blocked even if the LLM fails. Evidence is retrieved from local clinical guidelines via BM25 RAG.
- EHR agent: a local propose/approve layer turns the result into auditable write actions (
pending_human/blocked_by_safety) plus gaps vs prior visits. Approving persists the encounter locally. - UI: local web app (Python + Starlette) with a doctor-friendly flow: patient list, live listening, note review, human signature. Zero cloud calls.
Compliance by design
Local-only architecture aligned with ISO 27001 / ISO 27799 and Argentina's Law 25.326; human oversight in the loop (EU AI Act ready); every withheld order is traceable to its rule and evidence, a deterministic engine, not a black box.
https://prognosia.health/
Hacki