Respondr

Inspiration

Respondr was born out of a personal tragedy. A close friend of ours lost their father during an earthquake in California. In the chaos, emergency services were completely overwhelmed, and their call for help was never answered. That moment stuck with us — the idea that someone could be in desperate need, reach out, and still not be heard. We wanted to build something that ensures no emergency call ever gets ignored again. Respondr is our attempt to create a support system for dispatch centers during times of full capacity — a safety net that listens, understands, and prioritizes.

What it does

Respondr is an intelligent voice agent designed to automatically answer emergency calls when 911 lines are overloaded. It engages with callers in real time, gathering critical information such as the type of emergency, location, and the severity of the situation. Using an urgency ranking system powered by AI, it determines which calls need immediate escalation and which can be queued or flagged for later response. Respondr ensures that dispatchers are never blind to high-risk situations, even when human operators are fully occupied.

How we built it

For speech processing, we integrated Eleven Labs, which provided natural and responsive text-to-speech and speech-to-text capabilities — essential for clear, real-time interaction with callers. To give the voice agent a consistent, humanlike voice, we used a finetuned version of Open Sesame’s CSM-1B model on the LJ Speech dataset, allowing for nuanced and expressive text-to-speech performance. The backend was developed using FastAPI, enabling us to handle real-time asynchronous communication efficiently. All call data — including transcripts, urgency rankings, and metadata — is stored in a MySQL database.

Challenges we ran into

One major challenge was achieving real-time performance without sacrificing speech clarity or comprehension. Emergency calls involve panic, noise, and fast-paced speech, which makes accurate transcription difficult. We had to fine-tune latency and retry logic in our speech processing pipeline to ensure smooth interaction. Another difficulty was designing an urgency triage system that could distinguish between a true life-threatening event and a less critical situation. It took iteration, feedback, and model tuning to get it right. Lastly, without access to real emergency call datasets, we had to simulate realistic, high-pressure conversations.

Accomplishments that we're proud of

We’re proud of having built a working system that can autonomously manage emergency calls in overflow scenarios. Our voice agent is not only responsive and reliable but also emotionally sensitive — capable of engaging callers in high-stress moments without sounding robotic. We successfully integrated multiple AI components, built a full backend system, and developed a realistic simulation environment to test it all. The fact that Respondr can prioritize calls based on urgency and escalate life-threatening situations instantly is something we’re especially proud of.

What we learned

This project showed us just how powerful — and fragile — human-to-AI interaction can be in crisis settings. We learned the importance of tuning AI models not just for accuracy, but for empathy and speed. We gained deep experience working with speech technologies like Eleven Labs and realized the value of fine-tuning models on domain-specific data like emergency calls. FastAPI proved to be a fast and scalable framework for building the backend, and managing structured emergency data in MySQL gave us clarity and control. Most importantly, we learned that AI can augment human systems in ways that genuinely save lives — but only when designed with care.

What's next for Respondr

We plan to expand Respondr with multilingual support, advanced sentiment analysis, and a live dispatcher dashboard to display urgency scores, transcripts, and flagged calls in real time. We’re also exploring partnerships with city governments and emergency software vendors for pilot testing. A mobile version of Respondr is in the works, along with potential deployment in public venues like airports, malls, and schools. Our long-term goal is to make Respondr a nationwide safety layer — always listening, always triaging, always ready.

Built With

• csm-1b
• elevenlabs
• fastapi
• finetuning
• ljspeechdataset
• nextjs
• python
• tailwindcss