A hackathon where you build AI agents applied to real healthcare problems. You don't need to be a coder — the strongest coders wire the agent, the strongest thinkers design the prompt, the strongest communicators deliver the demo.
This is a resource kit, not a scaffold. Build however you want, with whatever tools you brought.
README.md ← you are here
Hackathon/
prompts.md ← the 3 challenge prompts in full
Workshop/
01_chat_only.py ← step 1: system prompt only
02_with_tool.py ← step 2: + tool definition
03_full_loop.py ← step 3: + agentic loop (full agent)
docs/
healthcare_primer.md ← what is MedEx, VBC, ED utilization, SDOH
data_dictionary.md ← what's in each table, key columns, gotchas
agent_building_guide.md ← the agent pattern: prompt + tools + loop
agent_design_framework.md ← 5 questions to answer BEFORE writing code
cloudflare_deploy.md ← 5-step Cloudflare Workers deploy guide
shepherd_system_prompt.md ← paste this into Claude Projects for a built-in teammate
data/
schema.sql ← D1 database schema (9 tables + patient_summary view)
*.csv ← all dataset CSVs if you prefer local files
examples/
python/agent_example.py ← full reference agent in Python (Anthropic SDK)
typescript/agent_example.ts ← same in TypeScript
Live services you can use today:
| Service | URL |
|---|---|
| Patient data API | https://uic-hackathon-data.christian-7f4.workers.dev/query |
| Hackathon guide chatbot | https://uic-hackathon-guide.christian-7f4.workers.dev/ |
| Patient lookup specialist | https://uic-patient-lookup.christian-7f4.workers.dev/lookup |
The patient dataset is hosted on a public read-only HTTP API. No account required — just make HTTP requests.
Endpoint: POST https://uic-hackathon-data.christian-7f4.workers.dev/query
# Try it right now
curl -X POST https://uic-hackathon-data.christian-7f4.workers.dev/query \
-H "Content-Type: application/json" \
-d '{"sql": "SELECT first, last, ed_inpatient_total_cost FROM patient_summary ORDER BY ed_inpatient_total_cost DESC LIMIT 5"}'Rules: Only SELECT statements are allowed. Responses are JSON { "results": [...] }.
Tables: patients, encounters, conditions, medications, observations, procedures, claims_transactions, careplans, patient_summary (pre-joined view — start here)
See docs/data_dictionary.md for what each table contains and data/schema.sql for the full schema.
117 synthetic patients (Synthea). Key facts to know for your agent:
| Stat | Value |
|---|---|
| Total healthcare costs | $27.9M across 8,316 encounters |
| Inpatient share | 35% of total cost, only 2.8% of encounters |
| Patients with at least one ED visit | 97 out of 117 (83%) |
| Top 3 patients by cost | Giovanni Paucek ($3.4M), Chad ($2.8M), Chantelle Oberbrunner ($2.5M) |
| Patients with no active care plan | 15 ED patients |
| Patients on opioids | 21 (19 are ED frequent flyers) |
| Patients with 5+ active medications | 25 |
| Patients with >$10K outstanding medical debt | 93 |
Pick one. 5 teams per prompt.
Build an agent that identifies patients at high risk of a preventable ED visit and drafts intervention recommendations for a care coordinator to review and approve.
Pattern: Filter → Score → Rank → Recommend → Human reviews
Build a conversational agent a care manager can interrogate to understand why a patient is expensive and which costs are reducible.
Pattern: Human asks → Agent queries → Presents findings → Human digs deeper
Build an agent that analyzes a patient's full record, identifies specific barriers (financial, social, logistical), and generates a barrier-informed care plan for a coordinator to review.
Pattern: Pull full profile → Identify barriers → Check care gaps → Generate plan → Human personalizes
Full prompts with data guidance in Hackathon/prompts.md.
Paste this URL into your AI agent (Claude Code, Cursor, Copilot, Codex — any of them):
https://raw.githubusercontent.com/csomora/INFORMS-UIC-Hackathon/main/SETUP.md
Your agent will check your environment, fix anything missing, and get you running your first query in under 10 minutes.
Option A — Cloudflare Workers (free, no API key needed)
Uses Cloudflare Workers AI — free on Cloudflare's free tier. No external API key required.
Fork this repo → scaffold agents-starter → connect to Cloudflare Builds → every push auto-deploys → demo a live URL.
See docs/cloudflare_deploy.md for the full walkthrough (~15 min setup).
Option B — Python (requires an LLM API key)
git clone https://github.com/csomora/INFORMS-UIC-Hackathon
cd INFORMS-UIC-Hackathon/examples/python
pip install -r requirements.txt
export ANTHROPIC_API_KEY=your_key_here # or OPENAI_API_KEY, GROQ_API_KEY, etc.
python agent_example.pyFree key option: Groq (OpenAI-compatible, generous free tier)
Option C — TypeScript/Node (requires an LLM API key)
cd examples/typescript
npm install
export ANTHROPIC_API_KEY=your_key_here
npm run startOption D — No code (Claude Projects fallback)
- Create a new Claude Project at claude.ai
- Paste
docs/shepherd_system_prompt.mdas the project instructions - Upload the CSV files from
data/as project knowledge - Claude becomes your agent — screen-record the conversation for your demo
| Criterion | Weight |
|---|---|
| Problem Framing — real, specific problem | 20% |
| Agent Design — multi-step, tool-using, goal-directed | 25% |
| Human-in-the-Loop — does human input meaningfully change the outcome? | 20% |
| Data Use — creative use of the dataset, not just loading it | 15% |
| Demo & Storytelling — problem → approach → demo → impact in 5 min | 20% |
| Patient | Why they're compelling |
|---|---|
| Giovanni Paucek | 63 ED/inpatient visits, $3.4M, 21 chronic conditions, overdose |
| Lindsay Brekke | 44 ED visits, chronic migraine, 10 conditions, NO active care plan |
| Chantelle Oberbrunner | 52 visits, $2.5M, 17 conditions, overdose |
| Soledad White | 35 chronic conditions (highest complexity), $276K ED/inpatient |
| Chad | 46 visits, $2.8M, 17 conditions, drug abuse |
Paste docs/shepherd_system_prompt.md into a Claude Project. Upload the CSVs. Ask it anything — it knows the dataset, the prompts, and the judging criteria. It will also tell you when you're over-scoping.