Larry

Inspiration

I'm a nontechnical founder building products for women in transition like divorce or the death of a spouse. I'm building Bea, an AI guide for the financial first year after loss, sand when I have been doing discovery calls with potential users, a really surprising and interesting thing came up. sometime with heartbreaking consequences. Many users told me the house was the thing nobody prepared her for. N The furnace made a new noise and she sat in the kitchen googling with wet eyes, because this was something her hubby always used to take care of.

And nationwide, there is a $98B repair backlog in our housing stock with older solo women disproportionately cost-burdened, and poor-condition homes sell 5–25% under market.

As a part of Bea, I've had it in the distant roadmap to build the home repair module, but when I saw the Build Week announcement, it felt like the right excuse to build the guy who answers the panic call.

What it does

Larry is a voice guide with retired-plumber energy. Two parts:

Safety Switches is a guided voice walkthrough of your own house. Larry directs, you find and photograph the important safety switch offs or key household systems: main water shutoff, breaker panel, gas meter (awareness only), furnace filter, etc.. Each stop is Find, Snap, Say, Test. Your photos and your own spoken descriptions become your House File, stored entirely on your device. That way, when you have an emergency, Larry has your back. He can tell you where the water main switch off is (because you located it together) and the house map builds the foundation for the next feature.

Then, the day the kitchen sink lets go, you call Larry. He asks one question at a time, slows down when you speed up, and remembers your house map: "Your main's behind the basement stairs, remember? We found it together." The saved photo appears on screen while he talks you to it. He stops hard at gas and electrical boundaries, preps what to tell the plumber ("supply line, not drain line, they'll know what to bring"), and names your win at the end. Because you did just handle it. And that type of capacity building is great for mental health and can contribute to a safer, and more valuable home.

How we built it

I'm nontechnical so I vibe coded it. I gave Codex the first brief to get the voice call functionality working and I was happy with what I saw. Codex used gpt-realtime-2.1 for the speech-to-speech work: turn-taking, interruption, Larry's spoken character, and function selection.

I used ChatGPT 5.6 for image gen to build Larry's persona and gave the image to Codex. Codex ran the build from the first implementation commit: turned the character spec into a constrained plan, verified the current Realtime/WebRTC function-calling contract, implemented the voice loop, the GPT-5.6 bridge, the local House File, and the photo-recall tools, and encoded the character, privacy, and safety invariants as 23 automated tests.

Challenges we ran into

There were a few bumps from a vibe coding perspective and things that I will want to continue developing. But nothing that is really notable. I've been vibe coding for a while with both codex and fable and other platforms and I'm used to having to go back and forth with the model. I think this was really close to the spec and needs limited fine tuning to be right on spec. I learned a lot of things that I will spec better and be more clear about in the future. Like for example which text is for the model to understand functionality and which text actually shows up on the user interface.

And overall for this specific audience, vulnerable women, who are learning to trust technology in a raw season of life, when I commercialize this, I will need to make sure there are clear privacy guardrails for her images and voice.

Accomplishments that we're proud of

I really love how Larry considers the house map when a user reports a problem. I think this is really cool functionality that I can build on in many applications and use cases.

What we learned

Oh oops covered this earlier. Vibe coding is always an exercise in learning to be more clear and more specific about your intended results and UX. However, the joy of vibe coding is also relaxing some of your PRD and seeing what the model comes up with. Sometimes it is shite and sometimes it is magic.

What's next for Larry

Larry grows up to be a module of Bea. Shared family access, maintenance rhythms, and a real handoff to human pros are all V2.

Built With

• 5.6
• codex
• gpt-realtime-2.1