Building CrisisTriageAI: Exploring the Boundary Between AI and Mental Health Support
I started this project because of a friend who works late-night shifts responding to suicide hotline calls. She often told me how exhausted the team is, how understaffed they are, and how difficult it is to juggle multiple high-urgency calls at once. I wasn't trying to replace human responders. Far from it. I wanted to explore whether AI could help triage calls in a way that supports them, not substitutes them.
The very first version of this idea was simple on paper: listen to multiple callers and rank them by urgency. Technically possible in theory, but ethically and legally impossible in the real world. Even as a student project, the boundaries were immediately clear. My friends in mental health told me: "I'm glad this is just research. AI in mental health is not something you deploy." They were right. So I reframed the project as a simulation: a safe environment to test how real-time AI pipelines behave when processing emotional speech.
On the engineering side, I wanted a modular pipeline that combined transcription, prosody features, and classification. The prosody extraction (pitch, energy, voice quality) was something I specifically pushed for, because urgency isn't just what people say but how they say it. Architecturally, the challenge wasn't writing code but drawing the right boundaries: deciding how audio moves through the system, where each stage runs, and how to surface results with low latency.
What I'm most proud of is the direction of the idea. Counterintuitive projects often lead to the clearest insights. Exploring ethically constrained spaces forces you to think harder about design, harm reduction, and responsibility. CrisisTriageAI is not a product and never will be, but building it taught me a lot about how AI systems behave when the stakes are high.
If I had more time, I'd work on deeper simulations and stress tests, exploring how multimodal signals interact under different conditions. The technical challenges matter, but so do the boundaries around them. This project was about learning both.