SummAID

Clinicians lose time reading long, multi-report patient histories, and cloud LLMs are off the table for protected health data.

A clinical assistant that reads a patient's report PDFs and writes a structured summary — medical journey, current status, action plan — where every claim links back to the exact source passage. It runs fully on-premise: a local Llama 3 through Ollama and an encrypted PostgreSQL store.

1. T+00:07 · STEP 01

   Parse

   PDFs are read with PyMuPDF (Tesseract OCR as fallback) and split into ~500-character chunks with overlap.

2. T+00:14 · STEP 02

   Encrypt & embed

   Each chunk is stored encrypted with pgcrypto beside a 768-dim nomic-embed-text vector in pgvector.

3. T+00:21 · STEP 03

   Hybrid retrieve

   Section regex (FINDINGS / IMPRESSION), vector cosine distance and keyword search run together — text is decrypted only inside SQL.

4. T+00:28 · STEP 04

   Continuity

   The previous summary and the doctor's edits are fed back as high-priority context, so summaries evolve instead of restarting.

5. T+00:35 · STEP 05

   Parallel generation

   The specialty is classified (universal, oncology, speech), then journey, status and plan are extracted concurrently, plus specialty-specific data.

6. T+00:42 · STEP 06

   Safety & render

   Deterministic temporal filters drop stale plan items, output is schema-validated, and the UI renders cards with clickable citations.

• Glass-box citations

  Every line of the summary opens the source PDF at the cited passage.

• Patient RAG chat

  Ask questions about a patient's history, with citations and speech-to-text input.

• Allergy safety gate

  A drug-specific allergy check blocks printing the digital prescription.

• Oncology & audiology views

  Tumour-size trends, TNM staging and biomarkers, plus audiogram charts.

• Doctor edits

  Append-only edits are merged at read time and carried into the next summary.