Clearline

Inspiration

The Palisades fire highlighted the lack of a centralized pipeline between live dispatchers that are receiving 911 calls and available data regarding wild-fire status from block to block. This causes evacuation assistance calls to be taken literally, and consequently send a team to evacuate a house that has already been cleared. We wanted to bridge this gap in order to make sure that resources are efficiently allocated, from dispatchers taking time to make reports that have already been filed to fire departments going to the same blocks more than once.

What it does

An AI agent is the first line that 911 goes to, in which it asks if the user or anyone that it knows within the Palisades area is in danger. If this is the case, or the user seems worried via sentimental analysis of keywords, the caller gets directly wired to a real 911 dispatcher. If it's about evacuation assistance calls, our agent will pull whatever house they are looking for from a database to see whether or not the houses' occupants still reside, or if they have chosen to evacuate. If the same address/person is reported more than once, the initial call's report is read to the caller, so they are aware of what's going on.

A dashboard then takes all of this information and marks houses green, yellow or red for successfully evacuated, currently evacuated or not evacuating - referencing the color tags officers place on home doors that wouldn't originally been able to be seen. There are two interfaces - one for dispatchers to see the current statuses of houses as they take calls, and one for the officers that are on site to update the statuses as they visit their assigned sites.

The dual-use aspect of our project is it's ability to help improve resource allocation during tragedies by routing transportation for aid to where they are needed most, instead of dead ends. On the flip side, its useful for ordinary life outside of a crisis, when you have not heard back from a vulnerable loved one, like a grandparent, or sick family member and need them to be checked in on.

How we built it

Our pipeline first has Retell AI handling the voice agent with a custom prompt and three tools - looking up an address, logging a call, and transferring a real dispatcher. We picked our model, __, by scoring candidates on scripted tests calls. The call server is a Node service on Railway. Every call runs through a set of named routing rules, and each decision is stored with the facts/rules that produced it, including but not limited to:

Speech slips in names and streets, implying stuttering and anxiety, and if detected hands the caller to a person mid-cal
A house's status is only shared with callers who give the resident's name
Reads the caller's words in case the agent mislabels a report Supabase holds 8,088 households built from OpenStreetMap building footprints (other lots maybe be available, but they are open as they have been the burn lots affected by the previous wildfires). Database triggers decide each house's color from both officer checks and phone reports. The dispatch board is a Next.js app with a MapLibre map that draws every building's real outline. Calls and color chang live in accordance to caller information. Officers get a phone map with GPS and taps that are saved offline in case there's no signal.

Challenges we ran into

We were told by mentors that are initial problem of being a general dispatcher helper was too broad, and would require us to take into account several factors. We were required to deepen our understanding of a specific niche in which our solution could be applied to - wildfires - and see how our solution can be integrated within the missing parts of the protocol.

Accomplishments that we're proud of

Our successful pivot and hyper fixating on a specific part of dispatch handling, and making sure that the buildings color change in real time after the agent is done calling a user.

Honestly getting support form professionals in the scene, like one firefighter/EMT that we talked to whom mentioned that our problem really could help save lives if it works to reduce the number of misallocated transportation resources. We had a very similar sentiment from a police officer we contacted as well.

What we learned

As 3/4ths of us had Dual-Hacks as a first hackathon, we were able to learn system design hands on, and setup frontend and backend environments such as Vercel for the frontend and Supabase for the backend. For non technical experience, we really did get an insight to how start ups make their design decisions, dissecting the problem before you solve it, considering funding, and even just how to gain credibility for our project. It was amazing talking to all of the founders and getting their support and insight on our project.