MedSift-AI

Inspiration

Healthcare professionals spend nearly as much time documenting patient visits as they do actually treating patients. Existing medical scribes often rely on cloud services, raising privacy concerns and limiting real-time usability. We wanted to build a solution that keeps sensitive data local while still providing powerful clinical intelligence.

MedSift-AI was inspired by the need for privacy-first, clinician-friendly AI that reduces burnout and improves patient care.

What it does

MedSift-AI is a fully local medical AI assistant that transforms doctor–patient conversations into structured clinical insights.

It:

• Transcribes visits using Whisper
  
• Automatically redacts PHI using Presidio
  
• Generates SOAP notes and care plans with local LLMs (Ollama + LLaMA 3.1)
  
• Produces patient-friendly summaries
  
• Matches relevant clinical trials and research
  
• Exports After Visit Summary PDFs
  

All processing happens on-device.

How we built it

We built MedSift-AI using:

• FastAPI backend
  
• Next.js frontend
  
• Whisper for transcription
  
• Ollama + LLaMA 3.1 for reasoning
  
• Presidio for PHI detection
  
• SQLite for visit storage
  
• External APIs (ClinicalTrials.gov, PubMed) for evidence lookup
  

The system follows a privacy-first pipeline: audio → transcription → PHI removal → structured extraction → research matching → PDF generation.

Challenges we ran into

• Running large models locally while keeping performance acceptable
  
• Handling audio transcription latency
  
• Resolving frontend-backend integration issues
  
• Managing Git merge conflicts during rapid collaboration
  
• Designing a clean medical UI under tight hackathon timelines
  

Accomplishments that we're proud of

• Built a complete end-to-end medical assistant in under 48 hours
  
• Achieved fully local inference (no cloud LLMs)
  
• Generated structured SOAP notes automatically
  
• Integrated clinical research matching
  
• Designed both clinician and patient-facing interfaces
  
• Implemented risk scoring and red-flag detection
  

What we learned

We learned how challenging real healthcare AI is — especially around privacy, UX, and performance. We gained hands-on experience deploying local LLMs, designing medical pipelines, and collaborating under pressure.

Most importantly, we learned how impactful privacy-first AI can be in sensitive domains like healthcare.

What's next for MedSift-AI

• Vector database integration for semantic medical search
  
• EHR system integration
  
• Mobile deployment
  
• ICD-10 coding support
  
• Fine-tuned medical LLMs
  
• Encrypted database storage
  

Built With

• clinicaltrials.gov-api
• fastapi
• github
• llama-3.1
• next.js
• ollama
• presidio
• pubmed-api
• python
• react
• sqlite
• streamlit
• tailwind-css
• typescript
• whisper-(openai)