DispatchAIBerkeley AI Hackathon · Grand Prize, 2024

Case study · June 2024

An empathetic AI that holds the line in the first ten seconds of an emergency.

DispatchAI answers 911 calls instantly, calibrates its tone to the caller's emotion, extracts location and severity in real time, and packages every incident for a human dispatcher in a live operator console. Built in 36 hours at the UC Berkeley AI Hackathon — and it won.

Open live console Tour the console

Award: Grand Prize; UC Berkeley AI Hackathon
Field: 1 / 293; of 930 builders
Outcome: $25K + Pad-13; SkyDeck Fund

Panel feedbackJune 23, 2024

Empathetic, technically deep, and the judges felt the demo could plausibly help real dispatchers tomorrow morning. That combination is what won the room.
Cal Hacks · Berkeley SkyDeck · Intel
UC Berkeley AI Hackathon panel

Built in

36 h

Stack lines

≈ 6.4k

Open source

Model + data

01The win

Grand Prize, Best Use of Intel AI — and a Pad-13 Golden Ticket.

At the UC Berkeley AI Hackathon (June 22–23, 2024), DispatchAI took the room. 930 builders, 293 submissions, organized by Cal Hacks and Berkeley SkyDeck.

Top placement

Grand Prize

$25,000 investment from the Berkeley SkyDeck Fund, $2,500 OpenAI credits, and a Pad-13 incubator Golden Ticket.

Track win

Best Use of Intel AI

Recognized for fine-tuning Mistral-7B on the Intel Dev Cloud and using IPEX to take inference from 2:53 to under 10 seconds.

Field size

930 builders

293 submissions over 36 hours at the UC Berkeley AI Hackathon, organized by Cal Hacks and Berkeley SkyDeck.

Post-hackathon

SkyDeck Pad-13

Golden Ticket admission to Berkeley SkyDeck's accelerator program, awarded as part of the grand prize package.

02The problem

The first ten seconds of a call decide everything — and the line is often empty.

Peak-load 911 centres routinely run minute-plus pickup queues. The agent in distress hears hold music. The dispatcher, two callers in, has no context. We wanted to fill that gap without removing the human from the chain.

60s+

Average 911 wait time during peak demand in major US cities, per dispatcher field reports.

50%

Of calls reportedly involve information that could be triaged or pre-filled before a human picks up.

Human dispatcher per call — protected. The AI fills wait time, never replaces final authority.

Real-time call cockpit

Incident queue, severity-coded map pins, live transcript, caller emotion, and street view assembled from a single FastAPI orchestrator.

Empathy as a first-class signal

Hume EVI emotion telemetry feeds the LLM context window so the agent's wording calibrates to the caller's distress, not just their words.

Human-in-the-loop by construction

Every action surfaces with provenance and confidence. Transfer is gated to a human dispatcher; the AI never closes the loop alone.

03The artifact

A 911 call answered, triaged, and handed to a human — in one minute.

Watch the submission video the team showed the panel. Same flow, same UI, same model — recorded the morning of judging.

DispatchAI product demo poster — The submission video shown to the panel: an end-to-end run from 911 inbound through Twilio → Retell → FastAPI → Mistral, surfacing a live transcript, emotion read, and dossier in the operator console.

04The system

Telephony, voice agent, emotion, inference, operator — all wired through one FastAPI orchestrator.

The voice loop is real-time. The model loop is async with confidence scores. The operator surface is a Next.js cockpit polling /api/calls every 5 seconds and merging server state with the active session.

System architectureTelephony · Voice agent · Inference · Operator

The caller reaches Twilio; Retell runs the voice agent over a websocket against the FastAPI orchestrator. FastAPI streams audio to Hume EVI for emotion, prompts the Mistral-7B LoRA on Intel Dev Cloud, and pushes geocoded incidents to the Next.js operator cockpit. A human dispatcher remains the final authority on dispatch.

05The model

A LoRA-tuned Mistral-7B, accelerated 10x on Intel Dev Cloud.

The team curated 911 call transcripts, fine-tuned with PEFT/LoRA, and ran inference on an Intel Data Center GPU Max 1100 using the Intel Extension for PyTorch. Both the model and a public snapshot of the dataset are open-sourced under MIT.

Latency · before / afterIntel Dev Cloud

Before

2:53

per response

→

After IPEX

<10s

same hardware tier

Reported in the team's submission: applying the Intel Extension for PyTorch to a Mistral-7B LoRA fine-tune collapsed per-response latency by an order of magnitude on the same Intel Data Center GPU Max 1100 tier — the work that earned Best Use of Intel AI.

Open-sourced

Both the fine-tuned model and a curated transcript snapshot are public.

Base model: Mistral-7B-v0.1
Tuning: LoRA · PEFT
Hardware: Intel Data Center GPU Max 1100
Optimization: IPEX (PyTorch)

Model · spikecodes/ai-911-operator Dataset · spikecodes/911-call-transcripts

06The interface

The operator console — deliberately dense, deliberately calm.

A non-interactive replica below renders against the same seeded incidents the live system uses for demo onboarding. Click an incident on the left to see the dossier, transcript and emotion update.

DispatchAI · Operator console

Read-only preview

Loading map…

FieldGolden Gate Bridge, San Francisco, CA

Selected incident

Earthquake Emergency at Golden Gate Bridge

37.8199, -122.4786

Live transcript

Caller × AI

AI operator connected

Retell · FastAPI orchestrator

Caller emotion

Fear35%
Confusion27%

Conversation

11 turns

Dispatch

911, what's the situation?

Caller

There's an earthquake — it's still shaking. Please, send help.

Dispatch

I hear you. You're doing the right thing by calling. Where are you right now?

Caller

I'm on the Golden Gate Bridge. The pedestrian walkway, north side. There are a lot of people down.

Dispatch

Got it — Golden Gate Bridge, north pedestrian walkway. I'm dispatching SFFD and CHP now. Roughly how many people are injured?

Caller

Maybe ten, fifteen. Some of them aren't moving. I don't know what to do.

Dispatch

Stay where you are if the structure feels stable. Do not try to move anyone who can't move on their own — you could make their injuries worse. Are you hurt?

Caller

I'm okay. I cut my hand on something but I'm okay. Just please tell them to hurry.

Dispatch

Units are en route. ETA four minutes. I'm staying on the line. If anyone near you is conscious and bleeding, apply firm pressure with whatever clean cloth you have. Can you do that?

Caller

Yeah. Yeah, I can do that.

Dispatch

Good. Keep talking to me. Tell me if anything around you changes — the bridge, the people, anything.

Human dispatcher remains the final authority

Read-only preview · sign in to open the live systemSeeded incident · CA22…0b3 · 2024-06-23

Want the real thing?

The live console wires up Clerk auth, the Retell voice loop, and Cloud Run polling.

Open live console

07The team

Four builders. Two days. One operator console.

Roles and contributions reproduced from the team's submission and follow-up posts. This portfolio belongs to one of the four — flagged below.

Human-AI interfaces · Backend · UX

Jasmine Wu

Started the project and solo-pitched the finalist demo to judges. Fine-tuned Mistral on real 911 transcripts, built the voice backend, and shaped the human-AI handoff working with real dispatchers.

LinkedIn · View profile

Machine learning · Backend

Spike O'Carroll

@spikecodes

Led ML and backend. Integrated Hume EVI for emotion, Twilio for telephony, and built the extraction + evaluation pipelines. Ran the LoRA fine-tune on Intel Dev Cloud and authored the open-sourced model and dataset on Hugging Face.

LinkedIn · View profile

Frontend · UX · Product

Kevin Wu

Owned the operator dashboard end-to-end. Built the real-time interactive cockpit in Next.js + TailwindCSS with a focus on calm, dispatcher-grade interactions under load.

LinkedIn · View profile

04You are here

Conversational AI · Voice agent

Bill Zhang

Built the conversational layer and the voice agent runtime. Integrated the LLM into the live call loop and stitched the real-time interactive cockpit you are reading this on.

LinkedIn · View profile

08Tradeoffs & honest limits

What this build is honest about — and what it isn't pretending to be.

Public-safety AI without an explicit limits page is a red flag. Here are the four most relevant ones.

Caveat

Dataset is small.

The public training snapshot has 518 transcripts. We ship the model and the dataset openly so it can be audited and grown — not as a finished product.

Caveat

Bias is real.

Accents, dialects, and cultural variation in distress aren't represented uniformly in 518 transcripts or in Hume's emotion model. Any production deployment requires demographic-stratified evals first.

Caveat

Humans dispatch.

The system is explicitly assist-only. Recommendations carry a confidence score; the dispatcher accepts, edits, or rejects. No outbound dispatch is initiated by the AI.

Caveat

Procurement is hard.

PSAPs need NENA / CJIS / SOC 2 alignment, integrations with legacy CAD and i3 NG911, and 6–18 month sales cycles. The hackathon build was deliberately scoped as a credible prototype, not a shipping product.

09Notes & sources

Every claim on this page is traceable to a public artifact.

GitHub, Devpost, Hugging Face, YouTube, Figma, and the original Vercel deployment. Recruiters — click through; nothing is hand-waved.

Notes & sourcesExternal

Try it

Step into the live console as a dispatcher.

Sign in (Clerk), add a phone number for the Retell voice agent to call, and the cockpit opens with the seeded incidents merged into live polling.

Open live console Watch the original demo

Sources

This portfolio cut maintained by Bill Zhang · Original build summer 2024