Bayesline

Filling details in the crucial gap between report and response for for first responders.

Comment

Inspiration

Recently, one of our team-mates got into first aid training and became quite interested in first responders on scene, specially in mass casuality situations. During our discussion of this problem, we were both quite interested in how much uncertainty first responders faced when approaching such situations, and how critical they can be.

What it does

The main aim of Bayesline was basically "How can we reduce uncertainty for the very first responder on the scene?" The way we approached it was by treating emergency calls not as sources of “facts,” but as streams of evidence. It listens to a live transcript of an emergency call, extracts key incident signals (location, incident type, hazards, possible casualties) in real time and tries to build an incident profile that increases certainty.

Instead of presenting definitive claims, Bayesline explicitly surfaces confidence and uncertainty, allowing responders and dispatchers to see what is most likely, what is still unclear, and how the situation is evolving minute by minute.

How we built it

We built Bayesline as a lightweight, simulation-friendly MVP focused on clarity and traceability.

  • A live voice stream representing an emergency call transcript.

  • Incoming voice chunks are processed to extract evidence such as hazards, location clues, and people estimates.

    • Different evidence is clustered together in a multi-modal information to determine a single incident state object.
    • Hopefully reduced risks of hallucinations by having LLMs being constrained by Bayesian, statistical, determinstic and other LLM as judge methods.
    • An append-only timeline logs every extracted claim, creating a full audit trail of how the incident profile evolved (Feat. Snowflake!)

NOTE: We hosted our project on the App Platform by Digital Ocean. However, due to some bug in DigitalOcean's billing platform, we couldn't get the credits, and hence couldn't fully deploy our project. Aside from this, we had our project entirely ready for deployment.

Challenges we ran into

A LOT OF BUGS OMG. Everything KEPT breaking and WOULD NOT stick together. Incorporating both of our techs together was a bit hard as well.

Accomplishments that we're proud of

Completing this project was really nice. One of us almost lost his voice by singing too much in the karaoke. We are also really proud of this idea and what it tries to solve for.

What we learned

DO NOT sing too much in karaoke. Also that perseverence really is key.

What's next for Bayesline

Honestly, theres so much. We can create more provably robust modelling techniques, better support for languages, better support for text during panic, standardised assement for allocation of resources etc.

Built With

Share this project:

Updates