Inspiration

In the "Golden Hour" of emergency response, connectivity is the first casualty. Disaster zones, remote rural areas, and subway systems often lack the internet access required for cloud-based AI tools. We wanted to build a solution that doesn't just work without the internet, but thrives there—bridging language barriers and providing expert medical reasoning when no doctor is available.

What it does

Sera is an autonomous, on-device AI medical liaison. It listens to voice input in multiple languages, translates it locally, and uses a reasoning agent (DeepSeek R1) to cross-reference symptoms with medical protocols. It provides instant, triaged advice to first responders—all with sub-200ms latency and zero data leaving the device.

How we built it

We architected Sera using the RunAnywhere SDK as our central control plane.

  • Voice Pipeline: We pipe raw audio through a local instance of Whisper V3 for real-time multilingual transcription.
  • The Brains: We run a 4-bit quantized version of DeepSeek-R1-Distill-Llama-8B directly on the device. This allows for complex "Chain of Thought" reasoning to verify symptoms.
  • Knowledge Base: A local Vector DB stores medical protocols that the agent can query offline.
  • Privacy: The policy engine is hardcoded to local_only, ensuring patient privacy by default. ## Challenges we ran into Running an 8B parameter model alongside Whisper on a mobile device was a memory challenge. We had to aggressively quantize the models and optimize the RunAnywhere SDK's policy router to ensure audio buffers didn't overflow while the LLM was "thinking." ## Accomplishments that we're proud of We achieved a true "offline-first" experience. Seeing the model correctly diagnose a complex scenario like High Altitude Pulmonary Edema (HAPE) in a simulated dead zone, with no server lag, was a breakthrough moment. ## What's next for Sera We plan to expand the local vector database to include more specialized trauma protocols and integrate with wearable sensors for direct vital sign monitoring without manual input.

Built With

  • android
  • deepseek-r1
  • ios
  • local-llm
  • offline-first
  • runanywhere-sdk
  • vector-db
  • whisper
Share this project:

Updates