Pulse

Logo
Base UI model we had in the beginning
Another base UI model we had in the beginning

Inspiration

We live in Hayward, and we kept watching great local spots lose regulars to chains without ever knowing why. Devout Coffee in Niles is a perfect example — two brothers who built something real, slowly losing customers to three Starbucks locations that opened within two miles. The tools that solve this problem cost $60,000 a year and were built for enterprise software companies, not coffee shop owners. We wanted to change that.

What it does

Pulse connects to a small business's Square or Stripe account and scores every customer by churn risk using their real transaction history. It shows the owner exactly who is slipping away, why the pattern looks the way it does, and what to do about it. From there, the owner can send a personalized email, get a phone script, or trigger a special offer — all written by AI and delivered as a voice note in their own brand's voice. Perplexity searches what competitors are trending in real time so the outreach is always relevant.

How we built it

We built Pulse as a single Vercel project with nine Node.js serverless functions powering a five-agent Claude pipeline that runs in parallel. Each agent has a specific job: scoring churn, recommending products, fetching live competitor prices via Perplexity, triggering surveys, and synthesizing everything into one personalized message. ElevenLabs converts that message into a voice note matched to the business's persona. The frontend was designed in v0 and built in React with Tailwind.

Challenges we ran into

Getting three AI providers to work together cleanly under a two-second latency target was harder than we expected. Perplexity doesn't always return clean structured data, so we had to build fallbacks that keep the pipeline running even when live prices aren't available. We also ran into browser autoplay restrictions with ElevenLabs audio that forced us to rethink how the voice feature works on different devices. A lot of the challenge was just deciding what to cut so the demo stayed tight.

Accomplishments that we're proud of

Honestly, we're most proud that the business model holds up when you push on it. We stress-tested every hard question before judges could ask it: the TAM, the competitive positioning, the moat, the pricing logic. On the technical side, running five Claude agents truly in parallel through a single orchestrator and getting a coherent output in under two seconds felt like a real win. And the ElevenLabs voice moment lands differently than any text demo we've seen at a hackathon.

What we learned

Each AI provider handles errors and edge cases completely differently, and you can't treat them as interchangeable when you're building a real pipeline. We also learned that the hardest part of a hackathon has nothing to do with code — it's deciding what not to build. Every hour spent on a feature that doesn't strengthen the demo is an hour taken from polish, and polish is what wins rooms.

What's next for Pulse

The two things we want to ship first are CSV upload for businesses not on Stripe or Square, and Google and Yelp review correlation so Pulse can tell the difference between a customer who left because of a bad experience versus one who just moved. Longer term, the feature we're most excited about is vertical benchmarking: showing a coffee shop owner how their retention compares to other coffee shops in their zip code. That's the kind of data no individual business can generate on their own, and it's what makes Pulse defensible over time.