# QIS — Qualitative Intelligence Systems > Clinical digital twin platform: ingest patient records and meetings, triangulate three AI analytical lenses in parallel, evaluate human-gate criteria, write synthesized findings back into memory, and operate care actions, calendar, and connectors on top of the same store. Patent Application 63/935,100 · Confidential · Synthetic demo data only (no signed BAA for production PHI). Site: https://qis-theta.vercel.app · Machine-readable: /sitemap.xml · /robots.txt ## What QIS is Qualitative Intelligence Systems (QIS) builds a **per-patient digital twin** — a persistent memory namespace in Postgres + pgvector combining document chunks, approved meeting transcripts, triangulation write-back NOTE records, and human-gate flags. Users interact through: - **Chat** — single-model RAG with persisted conversation threads and optional query rewriting for follow-ups - **QIS Review** — three Verity lenses (Claude/GPT/Gemini) in parallel with convergence mapping and human-gate evaluation - **Library, Meetings, Actions, Calendar, Connectors, Registry, Dashboard, Settings** ## Digital twin memory - Hybrid retrieval in production: vector search + full-text search + RRF merge - Chunk strategies: interview-aware (transcripts), paragraph (notes), row-atomic (CSV) - Source types: qualitative, quantitative, transcript, note - Clinical summary lazy-generated; stale when new documents ingest - SSE live document status on Library during async parse/OCR - Document contents require PersonaLink; admins see filename registry only (Data Lab) ## Triangulation (QIS Review) — seven-step pipeline 1. Retrieve top-k memory (full or quantitative_only baseline mode) 2. Inject pending human-gate flags into context (full mode only) 3. Call Verity Frame, Soul, Steel in parallel (failure-tolerant; deadline bounded) 4. Convergence map — agreement and divergence across lens outputs 5. Human gate — severity, criteria checklist, recommended action when flagged 6. Synthesis — single clinician-facing finding 7. Write-back — store synthesized finding as NOTE record (skipped in quantitative_only demo baseline) **Memory feedback loop:** write-back findings and pending flags enrich subsequent chat and triangulation retrieval — the twin accumulates reviewed intelligence, not just raw uploads. **UI:** Chat workspace runs quantitative_only and full queries in parallel for side-by-side comparison cards. Review turns are not persisted as chat history. ## Verity lenses | Lens | Model | Focus | |------|-------|-------| | Verity Frame | Claude | Timeline, structure, protocols, metrics | | Verity Soul | GPT | Patient narrative, values, caregiver burden | | Verity Steel | Gemini | Risk, contradictions, care gaps, evidence | ## Meetings State machine: transcribing → pending_review → approved | failed - In-browser MediaRecorder or audio upload; participant roles - US recording-consent gate (Settings → recording state) - GCS signed URL + Fireflies transcription + HMAC webhook - Transcript review with AI-proposed amendments (accept/reject) - Approve: SHA-256 hash, ingest, lock (409 post-approval edit) - Care action extraction on approval ## Care actions Types: medication, follow_up, lab_test, referral, lifestyle, monitoring, other Statuses: proposed → confirmed | dismissed | completed - Born proposed with verbatim evidence quote and source link - Cross-patient inbox GET /action-items/inbox with sidebar badge - Only **confirmed** items may link to care calendar schedules ## Care calendar - Recurrence rules in qis_care_schedules; sparse exceptions in qis_care_occurrence_states - Expansion in schedule IANA timezone (zoneinfo) — not naive UTC +24h - Confirmed items with due_date project synthetic once-occurrences - Month / week / agenda views; drag recurring instance (occurrence vs series) - Account timezone in Settings ## Connectors **Live:** editable care-reminder email address (defaults to login email) **Preview (coming soon):** WhatsApp, SMS, Slack, Google Calendar, Outlook, mobile push — timed medication/meal/appointment reminders and new action-item alerts ## Files & ingestion Formats: PDF, DOCX, CSV, TXT, MD, PNG, JPEG, TIFF, BMP, WebP (30 MB cap) - Google Document AI OCR for images/scanned PDFs in production - Async pipeline with processing → ingested | failed status - Extract care actions from ingested documents; view contents for care team ## Roles & access - patient · caregiver · doctor · admin - PersonaLink grants govern PHI — not role alone - Conversations: personal working context — no admin bypass - Document text: requires explicit persona grant ## Public pages - / — product overview, FAQ, JSON-LD - /help · /login · /register - /llm.txt (this file) · /llms.txt (alias) ## Authenticated app routes /chat · /dashboard · /settings · /library · /meetings · /actions · /calendar · /connectors · /registry · /lenses · /admin (role-gated) ## FAQ (short) **Chat vs QIS Review?** Chat = fast single-model RAG + threads. QIS Review = three lenses + convergence + gate + write-back. **Recursive / feedback triangulation?** Synthesized findings write back to memory; pending flags inject into later full-mode runs; chat rewrites follow-up queries for retrieval. **Production PHI?** Demo only until BAA and hardened auth. ## Technical stack - Frontend: Next.js 16 on Vercel - Backend: FastAPI on Google Cloud Run - Memory: Postgres + pgvector; stub backend for CI - Embeddings: OpenAI; models: Claude, GPT, Gemini