SymptomCheckerAi

Example UI vizualization on web browser (also scalable to work on mobile device)

Inspiration

When people feel unwell, the first question is “What should I do next?” Not everyone can interpret symptoms or decide between home care, a clinic, urgent care, or the ER. Symptom Checker AI was inspired by that gap: a fast, phone-friendly assistant that explains symptoms in plain language and suggests clear next steps, with a color-coded severity bar for instant understanding.

What I Learned

Streamlit for rapid, mobile-friendly UI in pure Python (tabs, popovers, custom CSS, session state).

OpenAI with Python for structured reasoning and JSON/JSON-Schema outputs I can reliably render.

Prompt design for safe, scoped triage (education, not diagnosis) and red-flag escalation.

Lightweight state management and UX patterns for small screens.

How I Built It

Design & Scope

Defined a system prompt: symptom-checker + triage assistant (not a doctor), with fixed triage levels: home_care, see_primary_care_1-3_days, urgent_care_today, emergency_now.

Chose a concise data contract for outputs (summary, differential, next steps, red flags, etc.).

Frontend (Streamlit)

Mobile-first “phone canvas” with custom CSS, a top navbar, and an avatar popover showing a human-readable profile card.

Tabs: Check Symptoms, History, About.

A green→red severity bar that positions a marker by triage level.

Backend / AI

Python + OpenAI (gpt-4o-mini) via Chat Completions.

Structured outputs with JSON Schema to keep the UI stable (and a JSON-mode fallback).

Minimal chat history via st.session_state to support follow-ups.

Severity Mapping (Tiny Math)

Challenges

Consistency of AI outputs: Early prototypes returned unstructured text. Solved by using JSON Schema responses and a fallback coercer for resilience.

Clinical safety boundaries: Balancing usefulness with safety. I enforced a narrow scope (education only), explicit red-flag escalation, and a persistent disclaimer.

Mobile UX polish: Streamlit is desktop-first by default, so I used custom CSS for a compact, app-like feel (larger tap targets, tight spacing, card styles).

Latency & reliability: Kept prompts concise, minimized history, and ensured an offline-demo fallback pattern if needed.

What’s Next

Editable multi-user profiles and consent flow.

Guideline retrieval from vetted sources, with citations.

Telehealth routing suggestions and localized care options.