I built and launched the beta of an iOS app that turns saved articles, PDFs, links, newsletters, and screenshots into private podcast episodes grounded in the original sources.

I consumed a lot of content but could not keep up with everything I saved. Links, articles, and references were scattered across Twitter, Slack, Instagram, my camera roll, and other places. Fio started from that frustration: it turns saved content into private episodes grounded in what the person chose.

1. 01

   Saved articles, links, newsletters, PDFs, and screenshots could become private audio that was easier to consume in the small gaps of a day.

2. 02

   Automatically generated podcasts delivered every week could turn occasional listening into a durable habit and give Fio a meaningful point of differentiation.

3. 03

   A source-grounded generation pipeline could produce episodes with enough quality and accuracy to be reliable, not just impressive in a demo.

A first AI prototype can be assembled in hours or a few days. The hard part was making a product that worked reliably: I built and tuned pipelines for AI agents around LLMs, retrieval, evals, prompt engineering, orchestration, and retry or re-evaluation decisions. I also had to connect APIs and integrations, keep terminology and numbers grounded in the source, and make TTS behave consistently. I defined and reached a 95% generation-success metric: 95% of podcast generations produced usable content while meeting the quality and accuracy bar, balancing reliability with the quality of the listening experience.

1. 01

   Designed the product experience from saved content to episode, including the home, player, and feedback flows.

2. 02

   Built the generation pipeline across summarization, script generation, and text-to-speech.

3. 03

   Used LLMs and AI agents as hands-on product and development partners, combining prompt engineering, evals, model iteration, and text-to-speech (TTS) to shape and debug the generation pipeline.

4. 04

   Created a landing page for the beta and built a waitlist that received a few dozen signups.

Building Fio taught me a lot about software: working with AI agents, AI agent pipelines, evals, prompt engineering, orchestration, APIs, integrations, and the full process of taking a product from a blank page to a finished experience across design, engineering, and launch. After building the waitlist, speaking with some users, and evaluating the idea, we learned that the problem was not especially urgent. Similar products already solved the same pain for free, and my differentiation hypothesis, automatically generated weekly podcasts, was not validated. The beta worked, but the product did not earn a strong reason to exist.