Vantura

The idea

I used to tutor, and one day I found myself talking with the parents of my students about their mornings. Most of them were slammed: getting ready for work, packing lunches, getting kids out the door. If they had interests outside of their jobs, there was rarely time left to keep up with any of it.

That conversation stuck with me. I kept thinking about what it would look like to hand someone a briefing built just for them, something they could read over coffee or listen to in the car, and walk away caught up on everything they cared about. Around the same time, a few of us were watching the new apps built on top of the iMessage API, things like Poke and Tomo, and it clicked that this could be the right home for the idea. If Vantura lived right inside iMessage, it wouldn't feel like another app competing for your attention. It would just feel like a text from someone who already knew what you cared about.

How it works

The first thing a new user does is go through onboarding (there's a clip of it in the video above), where we learn what they actually want to follow: specific topics, sports teams, stocks, whatever matters to them.

From there, we go pull the news. We use a mix of News APIs and other sources, along with Grok, which we leaned on to search by topic. It wasn't the most efficient approach. We hadn't built a clean filter yet to separate the useful articles from the noise in the News API results, so in the beta we just let Grok do that work for us, and it held up well enough.

Once we have a day's worth of material for someone, we turn it into a set of visual news cards they can scroll through and react to right inside the thread, and a podcast style script that gets sent over to ElevenLabs, which is what actually produces the audio briefing that lands in their inbox each morning.

What we learned

The hardest part of all of this had nothing to do with news or audio. It was getting a language model to behave. Model output is stochastic by nature, so pinning down something consistent and safe enough to send to real people took a lot of trial and error. I spent a long stretch deep in prompt engineering, fine tuning, and RAG before we found something that worked: a loop with actual guardrails built in, regex checks plus an LLM acting as a judge over its own output (we used DeepSeek for that since it was cheap). If something slipped through wrong, the judge would catch it before a user ever saw it. Getting Vantura to call the right tools at the right time was its own version of that same fight, since being right most of the time isn't good enough when someone is trusting you with their morning.

The rest of what I learned had less to do with code and more to do with running something like a business. The iMessage API we depended on wasn't cheap, and it showed up fast in our burn rate. Asking people to pay for Vantura turned out to be its own challenge, since the briefing alone didn't always feel like enough value to justify a subscription. We ended up around a hundred users. I eventually stepped back from the day to day to go build other things, but looking back, it was one of the more fun projects I've worked on.

Along the way we picked up some real signs that we were onto something. Founder University accepted an earlier version of Vantura into their pre seed program, and even though the product had changed a lot by the time we actually shipped, that early validation helped us sharpen our pitch and get comfortable in investor conversations. ElevenLabs also gave us about five thousand dollars in credits, which is really what made the daily podcast possible in the first place.

What's next

These days I'm spending more time exploring agent harnesses, LangChain in particular, and thinking about what a next version of something like Vantura could look like. Wherever it goes from here, I'm glad I built it, and I hope it keeps finding its footing.