Siren

Our Inspiration

Our inspiration came from a heartbreaking experience involving our teammate Chelsi. Her grandmother was experiencing a medical emergency but chose not to call an ambulance because she feared the cost. Instead, she went to the hospital herself, where she tragically passed away while waiting in the emergency room.

Her story made us realize something: Why should the quality of emergency care before hospital admission depend on whether someone arrives in an ambulance?

Ambulances provide more than transportation; they also coordinate medical help. Paramedics assess patients, communicate with emergency departments, identify appropriate hospitals, and give medical teams time to prepare before arrival. But patients who arrive independently often have none of these advantages.

We wanted to change that.

We built Siren to help bridge the gap between patients, paramedics, and hospitals. Using AI-powered triage, intelligent hospital matching, and real-time communication, Siren aims to bring critical pre-arrival coordination to people who might otherwise enter an emergency department without advance notice.

Because nobody should lose a loved one simply because they were afraid of the cost of getting help.

What It Does

Siren turns your car into an ambulance.

Patients describe their symptoms in five languages, and Claude uses AI-powered triage to assess their urgency. Siren then recommends an appropriate emergency room based on location, hospital capacity, and specialist availability. For life-threatening emergencies, patients are instructed to call 911.

Unlike traditional navigation apps, Siren also prepares hospitals before patients arrive. A live dashboard provides medical teams with incoming patient information, symptoms, and estimated arrival times. Our AI-powered Charge-Nurse Assistant helps coordinate resources and bed assignments with nurse oversight.

Siren also connects ambulance crews with hospitals, provides voice-based patient check-ins, and allows families to track their loved ones without revealing sensitive medical information.

Our mission: Get patients to the right ER, with the right team ready before they arrive.

How We Built It

We built Siren using React for the frontend, FastAPI for the backend, and SpacetimeDB to power real-time communication between patients, paramedics, and hospitals. SpacetimeDB allows patient information, hospital notifications, and emergency status updates to synchronize instantly across interfaces, while enforcing access controls and safety rules at the database level.

We integrated Claude for AI-powered medical triage and our Charge-Nurse Assistant agent, alongside Cartesia Ink and Sonic for multilingual speech-to-text and text-to-speech. We also used OSRM for route calculations and Apify to collect publicly accessible hospital information, incorporating California HCAI and CDPH datasets to inform hospital recommendations.

To evaluate our AI triage system, we used MIMIC-IV-ED, comparing its predictions against real nurse-assigned triage decisions.

Challenges We Faced

Accessing Real-Time Hospital Data

One of our biggest challenges was obtaining reliable, real-time information about hospital capacity, available beds, medical specialists, and equipment. Unlike restaurant availability or traffic conditions, this information isn't readily accessible through public APIs.

To overcome this, we used Apify to collect publicly accessible hospital information and SpacetimeDB to synchronize updates across our platform in real time. For operational information that wasn't publicly available, we used simulated data and clearly distinguished it from verified hospital information.

Preventing Speech Recognition Errors

During testing, we discovered that our speech-to-text system occasionally generated sentences from background noise, triggering unnecessary notifications to nurses. We implemented additional filtering to prevent unreliable transcripts from activating downstream workflows.

Enforcing AI Safety Constraints

Our Charge-Nurse Assistant initially needed stronger restrictions on bed assignments. We introduced database-enforced safety rules to ensure critically ill patients could only be assigned to Resuscitation or Acute care beds, while requiring nurse approval for significant decisions.

What We Learned

Our biggest lesson was that AI should be measured, not just demonstrated.

Our initial Claude triage prompt performed well in simple demonstrations but struggled with borderline high-risk patients. We evaluated our system against 206 real nurse-assigned triage decisions from MIMIC-IV-ED and refined our approach, improving high-risk detection from 84% to 96% in our testing.

We also learned that integrating AI into healthcare requires more than an accurate model. Reliable data, real-time communication, privacy protections, and human oversight are equally important.

Most importantly, we realized that the nearest emergency room isn't always the right emergency room. What matters is reaching a hospital equipped to provide the care a patient needs, while giving medical teams time to prepare.

Siren's goal is to make that coordination possible before patients arrive.