Lysn.to: Emails and articles turned into a personal podcast

We were one of three companies on the initial call. The others brought proposals; we brought a working proof of concept and a list of the specific points worth solving before money went into full development. That won us the project, and it set the tone for the whole engagement: a partnership rather than a delivery contract, with paid MVP development and future equity on the table.

The build followed four rules. Keep the stack lean - the minimum of infrastructure, the maximum of speed. Accelerate development with AI tools. Ship to production early and collect real feedback instead of polishing in a drawer. And validate before monetizing - we build only what is needed to learn.

We defined the product in January 2026 and had it running in production at lysn.to in March 2026. An alpha test with the first users started right after launch.

The hard question was never whether the product could be built. It was decisiveness: what to build it from, so that the economics and the audio quality both made sense. Every AI building block has several strong candidates, and it is easy to sink weeks into evaluation. We time-boxed the key decisions to days, made them with data, and moved on.

The MVP deliberately ships without a payment gateway: the goal is to validate the product with the first 100 active users before monetization. Then come the business plan and fundraising. Together with the founders we are also shaping the long-term form of the partnership, including a possible Cookielab stake in the project.

Lysn.to is a snapshot of what an AI-native build looks like in 2026: AI at the core of the product and inside the development process itself, a lean stack chosen for speed, and a deploy to production as early as the product allowed.

The product is a pipeline. A user signs up with a magic link, no password, and gets a unique email address. Anything forwarded there goes through an LLM - GPT-4 or Claude - which cleans the text and rewrites it for listening. That step separates Lysn.to from ordinary text-to-speech tools, which read word for word. The finished audio publishes as a private RSS feed the user adds to Spotify or Apple Podcasts, next to their regular shows.

The stack is deliberately minimal: Supabase covers the database, hosting and asynchronous processing through Postgres message queues, Cloudflare handles DNS, and almost nothing else sits around it. Claude wrote a large share of the code; our job was direction, architecture and quality - generated code needs the same scrutiny as human code, sometimes more.

Most of the engineering energy went into choosing the text-to-speech engine. We tested three models on quality, speed, and price per minute of audio. Cartesia Sonic 3 was very fast and cheap at roughly $0.034 per minute, but Czech sounded robotic. ElevenLabs had, to our ears, the best voice quality and a conversational mode, at higher cost and slower generation. Google Chirp 3 HD matched that quality but was slower and tied to GCP. We deployed a combination of Cartesia and ElevenLabs - and gave the whole evaluation days, not weeks.

The real world added its own homework, like Outlook opening magic links automatically before users could. We built a workaround and moved on.

Feedback from the first users arrives through a button in the app that posts straight into Slack, and real edge cases get fixed as they show up. Once validation with the first 100 active users is done, the next steps are monetization and deepening the processing pipeline.