Crohns Care

The problem

Crohn’s symptoms are often recorded as messy daily notes: slang, skipped days, one-word entries, and out-of-order fragments. That makes it hard for patients to communicate how they have actually been, and hard for a clinician to scan a diary into a clear trajectory.

How it works

Choose a synthetic diary or paste your own. A language model extracts Harvey-Bradshaw Index fields from each day (wellbeing, pain, liquid stools, extra-intestinal signs, medication adherence). Python then scores HBI, baseline vs current windows, and adherence — the model does not do the arithmetic. A short outpatient brief and PDF are generated from those metrics plus retrieved log excerpts.

RAG

Each run builds an ephemeral Chroma index of that diary only. The synthesizer does not read every day in full. It retrieves a few temporally labelled chunks for dietary triggers, complications, medication mentions, and stool pattern, then writes the narrative from those spans plus the deterministic stats.

Guardrails

• Structured extraction via a Pydantic schema, not free-form diagnosis.
• Only Bristol 6–7 stools count toward liquid stool score.
• Dietary links are only flagged inside a 24–48 hour window, and only if they appear in retrieved notes.
• The brief is non-prescriptive: no drug, diet, or referral advice.
• Diaries stay in memory for the request; collections are not shared across patients.

Source code on GitHub

Disclaimer. This is not a medical diagnostics tool. It is non-diagnostic decision support for demonstration. It does not diagnose Crohn’s disease, assess flare risk, or replace a qualified clinician. Do not use it for treatment decisions.