💡 Inspiration

Healthcare doesn’t fail inside hospitals — it fails after patients leave.

We were struck by how patients are expected to manage medications, symptoms, and follow-ups on their own, using documents they barely understand. Research shows that patients forget up to 40–80% of what doctors tell them, and nearly 1 in 5 patients experiences harm after discharge, much of it preventable.

We realized the problem isn’t just medical — it’s about understanding, memory, and access to guidance.

That’s what inspired us to build AfterCare.

🚀 What it does

AfterCare turns medical documents into an interactive, patient-friendly experience.

Upload a discharge summary, prescription, or report
The system extracts structured medical information
Patients can ask questions in natural language
The system responds with clear, grounded answers based on their own documents

It effectively acts as a post-discharge copilot, helping patients:

Understand medications
Interpret test results
Follow instructions correctly
Know when to seek help

🛠️ How we built it

We designed a lightweight, end-to-end pipeline:

Document Processing

Extract text using PyMuPDF
Convert unstructured medical text → structured JSON using Mistral

Retrieval (RAG)

Chunk documents into segments
Generate embeddings using Mistral
Retrieve relevant context using cosine similarity

Q&A System

Use Mistral LLM for grounded responses
Combine:
- Retrieved document chunks
- Structured patient summary
- Optional web snippets (DuckDuckGo API)

System Architecture

Frontend: Next.js + Tailwind
Backend: FastAPI
Storage: In-memory session store (no DB for MVP)

The result is a fast, hackathon-friendly system that prioritizes clarity and reliability over complexity.

⚠️ Challenges we ran into

Messy medical documents: Real discharge summaries are inconsistent and hard to parse reliably
Balancing simplicity vs accuracy: Keeping answers understandable without losing correctness
Grounding responses: Preventing hallucinations while still answering naturally
Time constraints: Building extraction + RAG + UI in a short hackathon window
Scope control: Avoiding overbuilding (EHR integration, authentication, etc.)

🏆 Accomplishments that we're proud of

Built a fully working end-to-end system in hackathon time
Successfully converted unstructured medical text into structured patient summaries
Implemented a RAG pipeline grounded in patient-specific data
Designed a system that prioritizes safety and non-hallucination
Created a clean, demo-ready UI that clearly communicates value

Most importantly, we didn’t just build a chatbot —
we built something that addresses a real, high-impact healthcare gap.

📚 What we learned

In healthcare, clarity matters more than intelligence
RAG is only useful if the input data is clean and structured
Users trust systems that say “I don’t know” instead of guessing
The hardest part is not answering questions —
it’s understanding messy real-world data