OushodhSathi

A pharmacist for every elderly home in Bangladesh: we scan every medicine a parent takes, catch dangerous duplicates and interactions, and a clinician confirms every finding before the family sees it.

Live demo: link added after deploy (how) · Docs: the /docs page of the app · Demo video: YouTube link added after upload · Reviewer login: the dashboard at /review asks for a name and the REVIEWER_TOKEN from backend/.env (the quick start sets it to demo-reviewer)

---

The problem

• Hypertension is now the most common illness in Bangladesh, at 78.28 per 1,000 people (BBS Health and Morbidity Status Survey 2025, reported by BSS and The Daily Star).
• Only 11% of people with hypertension had their blood pressure controlled below 140/90, a background figure in the WHO-HEARTS evaluation in Bangladesh (Heart, 2024).
• 69% of geriatric outpatients were prescribed five or more medicines in a small 2011 study of 35 patients at a private hospital (Trends of prescribing for out-door geriatric patients).

An elderly patient often collects medicines from a hospital outpatient doctor, a village doctor and the local drug shop. Each sees only their own prescription, and the same molecule is sold under many brand names, so nobody notices that Napa and Ace are both paracetamol, or that two doctors have each started amlodipine.

How it works

flowchart TD
    A["Family photographs<br/>every strip, box<br/>and prescription"] --> B["Vision model reads<br/>brand, strength,<br/>expiry, confidence"]
    B --> C{"Brand known?"}
    C -- "yes" --> D["Brand → generic<br/>(exact, then strict fuzzy)"]
    C -- "no" --> W["Web lookup on Bangladeshi<br/>medicine sites (Tavily / Perplexity)<br/>must cite a source → unverified"]
    W --> D
    D --> E["Rules engine<br/>duplicates · same class · interactions<br/>Beers 65+ · expiry<br/><b>no AI</b>"]
    E --> F["Pharmacist / doctor review<br/>confirm or dismiss, with a note"]
    F --> G["Family view in plain Bangla<br/>+ weekly report<br/>+ medicine passport (QR)"]

Anything read with under 0.75 confidence, and every web-found brand match, is shown to a human for confirmation before it enters the list.

Screenshots

All screenshots use the invented demo household of Rahima Begum, 68, Cumilla.

Why it's safe

AI for perception, rules and humans for judgement. Models read photos and write explanations; they never decide what is dangerous.

1. Rules make every finding. Safety flags come only from the deterministic engine in backend/app/rules.py. Language models never create, remove or change the severity of a flag.
2. A clinician confirms before the family sees anything. Flags start as pending_review. The family view of the API returns only confirmed flags, enforced in the API, not just the UI, and covered by a test.
3. No instruction to stop or change a medicine. No text tells a patient to stop, start or change a medicine or dose. Explanations are checked for doses and stop-words and end with a fixed safety sentence.
4. Uncertain readings are marked. Extractions under 0.75 confidence and web-found brand matches are always shown as needing human confirmation.
5. Names stay out of models. Only medicine photos and medicine names are sent to a model, never patient names or phone numbers.
6. Photos are not kept. Photos are processed in memory and discarded after extraction.

These rules live in AGENTS.md and bind every coding agent that works on this repository.

Architecture

flowchart TB
    UI["Next.js 15 frontend<br/>family app (Bangla) · reviewer dashboard · passport · /docs · /demo"]
    API["FastAPI backend<br/>family view returns confirmed flags only"]
    AG["Shelf agent<br/>Pydantic-AI, typed output"]
    R["LLM router<br/>cache + fall-through"]
    P1["OpenRouter<br/>(free models, auto-picked)"]
    P2["NARAROUTER → EXPLAB → AGENTROUTER<br/>(OpenAI-compatible fallbacks)"]
    P3["Ollama qwen2.5:3b<br/>(local, Bangla explanations first)"]
    BM["Brand matcher<br/>exact → fuzzy ≥ 88"]
    WL["Tavily / Perplexity<br/>unknown-brand lookup"]
    RE["Rules engine<br/>deterministic, no AI"]
    DB[("SQLite<br/>items · flags · reviews · cache")]
    N8N["n8n automations<br/>weekly report · doctor alert · refill reminder"]
    MCP["MCP server<br/>read-only tools"]

    UI -->|HTTPS, JSON| API
    API --> AG --> R
    R --> P1 & P2 & P3
    API --> BM --> WL
    API --> RE
    API --> DB
    API -->|red flag webhook| N8N
    N8N -->|GET report| API
    MCP -.->|calls the same modules| BM & RE

SQLite is enough for the prototype and lives on a persistent disk when deployed (render.yaml).

This is the beginning of the README. Read the rest at the original. Read the full README ↗