Hearth AI's Ashe Magalhaes on Building With Your

There is a particular kind of founder who arrives at their thesis not through a market gap analysis but through something closer to a reckoning. Ashe Magalhaes, founder of Hearth AI, describes sitting in a neuroscience class at Stanford when a guest was brought in—Paul Kalanithi, the neurosurgeon who would go on to write When Breath Becomes Air—and being told: this person has a terminal illness, ask them anything. The experience lodged itself somewhere important.

"The satellites are amazing, the solar cars are amazing, all the ML stuff is cool," Magalhaes told Romain Huet, OpenAI's Head of Developer Experience, in a recent episode of the Builders Unscripted series. "But I think what I'm going to care about is how connected I was to myself and to other people in my life, and how much presence and love I exchanged."

That is not a typical founding story. It is, however, a coherent one—which is more than you can say for most.

Magalhaes's technical biography reads like a deliberate tour of frontier systems. Stanford solar car team, where she spent two years on telemetry and CAD modeling before driving the finished vehicle across the Australian Outback in the Bridgestone World Solar Challenge—five days, Darwin to Adelaide, at 90 kilometers per hour in a car her team built entirely themselves. Then satellites at NASA. ML engineering at Airbnb. Then, in 2022, Hearth AI, which she describes as the first agentic CRM.

That last claim deserves some scrutiny, if only because "first" is a word founders deploy enthusiastically. But the timeline holds up. Hearth launched in 2022, was running agentic workflows on GPT-3.5 in 2023, and was attempting structured outputs via JSON schemas before OpenAI had even shipped that feature formally. Huet, who was at OpenAI during that period, confirms it: "You were trying to do retrieval with JSON schemas. You were so early."

What Magalhaes was building, beneath the technical scaffolding, was an answer to a problem she frames as almost civilizationally basic: modern humans are connected to more people than ever before and feel lonelier for it. The agentic CRM was one attempt at that problem—a multiplayer pass, she calls it. Her current work on what she terms "relational intelligence" is another iteration of the same question.

The thesis: AI should extend the brain's ability to reason about who we're connected to and why. Not replace human relationships. Augment the cognitive infrastructure that relationships run on.

Whether that thesis will prove out is genuinely open. The history of tech products built around connection is not uniformly encouraging. Social networks were also, in their founding rhetoric, about human connection. The gap between what a product is designed to do and what it actually does to human behavior is wide enough to drive several cautionary tales through.

Magalhaes seems aware of this tension—her whole emphasis on presence, on reducing screen time, on a "second brain" that works for you rather than demanding your attention, reads as a deliberate departure from the engagement-maximizing playbook. She is building a wearable device alongside her software work, and her articulated vision for it is notably careful: no ambient recording without consent, social etiquette protocols for what gets captured and what doesn't, a gesture-based system where the person across from you can signal comfort or discomfort with information being logged.

"I don't like ambient recording," she said plainly. "I do think we will need a stream of consciousness that is systematized for our second brain. But first I want to play with what might the future social etiquette be in a way that feels elegant and not gauche."

That is a reasonable instinct. It is also, it should be noted, exactly the kind of instinct that tends to get rationalized away when products meet growth pressure. The question of who controls consent norms in ambient intelligence systems is not a small one—it is precisely the kind of question that regulators in Europe are already grappling with under the AI Act, and that U.S. policymakers have so far mostly declined to engage with seriously. Magalhaes is thinking about it at the product design level. Whether the broader industry will is a different matter.

The operational side of how Magalhaes actually builds things is where the episode gets most concrete, and most interesting as a window into how serious solo developers are working right now.

Her setup: a public-facing personal site, ashe.ai, backed by a private "secrets page" that functions as a prototype incubator. New ideas get built there using OpenAI's Codex. They get posted publicly. If the signal from other builders is strong enough—comments, requests to open-source, engagement—she instructs Codex to extract the prototype into a standalone product. If not, it stays in the sandbox.

The instrumentation layer is not an afterthought. Every product she ships gets observability tooling from day one, routed through a Slack workspace where agentic workflows live in dedicated channels. News. Birthdays. Goals. Money. The agents surface alerts, catch errors, manage tasks. She demonstrated this live during the conversation—there was a bug in aesthetic.video, her video collaboration platform, and she simply instructed the agent to fix it while the interview continued.

"I showed someone at a bar once and they were like, are you the only person on this Slack workspace?" she recounted. "I was like, yes. And they started laughing. But it's the future."

Possibly. The architecture she is describing—agentic workflows as personal operating infrastructure, with a human as the creative director rather than the executor—is something a lot of people are experimenting with right now, with varying degrees of rigor. What distinguishes Magalhaes's version is the insistence on a reliable foundation rather than what she calls "the fragility of a vibe coded product." Speed matters, but not at the cost of something you'd actually use every day.

The aesthetic dimension is not incidental to her. She talks about beauty with the same register she uses for technical architecture. She wants to interact with her tools the way one might interact with something well-designed—not just functional, but resonant. This is not standard engineering discourse, and I find it worth taking seriously rather than dismissing as founder romanticism. The products people actually adopt tend to be the ones that feel good to use. That is not a soft criterion.

Magalhaes draws a distinction that I think is genuinely useful for understanding this particular moment in AI development: the difference between building products and building systems. Earlier eras of software development had you choose one or the other. What she is describing—and what tools like Codex are making more tractable—is maintaining both simultaneously. The prototype is the system is the product, iterated in public, shaped by real feedback, extracted when validated.

"The name of the game is speed," she said. "But what will you build that you yourself use every day that you find to be a beautiful and elegant experience, and that you can surface to users in a way that is sustainable?"

The self-use criterion is worth highlighting. It functions as a filter that is easy to articulate and surprisingly hard to fake—you either reach for the tool or you don't. In a moment when the barrier to shipping software has dropped far enough that nearly anything can be built, the question of what is worth building becomes more important, not less. Magalhaes's answer—build what you actually need, make it beautiful, share it, watch what happens—is not a framework so much as a practice.

Whether "relational intelligence" becomes a product category, a company, or a wearable that changes how people move through conversations with other humans is genuinely unknowable from here. What is clear is that the underlying problem Magalhaes has identified—that we have more connections and less presence than ever before, and that this is a design failure worth engineering against—is real. The question is whether AI is the right instrument for it, or whether it is yet another technology that promises connection and delivers something more complicated.

By Samira Okonkwo-Barnes