Modo Matcha Sales Scout

Inspiration

This is a real business I'm close to. Mōdō Matcha has booked twelve brand partners in under a year. Meta, Adobe, Benefit Cosmetics, Princess Polly, SEGA. That sounds great, but almost all of it is inbound. People find them first. Word of mouth, referrals, agencies that already know the brand. There's no real outbound today. Not because it wouldn't work. Because a one person shop doing sales, production, and delivery has never had a spare hour to build one. That's exactly the kind of problem this hackathon asked for. Routine. Repetitive. And the only real value shows up in the last five percent, deciding if a lead is even worth chasing.

What it does

Give it nothing but a prompt to go find work. The agent searches the live web for a real company in LA or Orange County that just showed a signal, like a store opening, a funding round, or a new product launch. It won't move forward without a real link backing it up. Then a second pass scores how good a fit that company is, based on Mōdō Matcha's real rules: service lines, how much lead time we need, past case studies, the client list. Then it writes an email that names the exact signal and picks the one case study that actually matches. It found and pitched Erewhon's new Orange County location and MOTHER Denim's Beverly Hills store opening. Both real. Both with links you can click. Nothing made up.

How we built it

It runs on the Strands Agents SDK, using Amazon Bedrock for the model, with Tavily doing the live search. It's two agents, not one. The first agent has the search tool and just writes up what it finds in plain text. The second agent has no tools. It takes that text and turns it into a real structured object: company, signal, score, reasons, and the email. We split it that way because Strands doesn't like using a tool and doing structured output in the same call. They fight each other. Splitting them fixed the bug. And honestly, it's just a better way to build it anyway.

Challenges we ran into

The original plan was five agents. Scout, Researcher, Scorer, Writer, Critic. All chained together, deployed on Bedrock AgentCore, with DynamoDB storing leads, AgentCore Memory, and a dashboard where a human approves every email. That's still the full plan. It's written out in SPEC.md. But we only had a few hours to actually build something. So we cut it down to two agents. That's the hardest part of the idea anyway: can an AI actually find a real company, check it's real, and write a good pitch. No deployment, no dashboard, just the core loop.

The best bug of the whole project: the .env file loaded in the wrong order, so the very first real run had no search key at all. Instead of making up a fake company, the agent just said it couldn't find one. That's not a bug. That's the agent doing exactly what it's supposed to do. It just happened by accident before we meant it to.

Accomplishments that we're proud of

Everything shown in this submission is real. Real search results, real Bedrock calls, real companies, real links a judge can click and check themselves. Nothing was faked to make the demo look better than the code actually is. The pitch also got more honest as we built it. At first we said the business owner was too busy to do outreach. The real story, confirmed by the owner, is that outreach never existed at all. That's a better story. And it's true. So that's the one we're telling.

What we learned

We learned that Strands structured output and tool use don't mix well in one call, and splitting them fixes it. We learned that Bedrock model access depends on your account, so it's worth testing with one simple call before building anything on top of it. And we learned that the most convincing line in a pitch is sometimes the one you almost don't say. Saying this business has never had an outbound channel at all was stronger than any efficiency pitch could have been.

What's next

The full system is in SPEC.md. Five agents working in a set order. Deployed on Bedrock AgentCore Runtime. Triggered on a schedule with EventBridge. DynamoDB tracking every lead so nothing gets contacted twice. AgentCore Memory so the scoring gets smarter every time the owner approves or rejects something. And a dashboard so no email ever goes out without someone actually saying yes.

Built With

• amazon-bedrock
• anthropic-claude
• pydantic
• python
• strands-agents-sdk
• tavily