Alex

I'm building @mato_hq, AI agents that interview humans, starting with podcasts. $0K to $18.5K MRR in 3 months; nobody has churned; about 30 shows running. one of them sits at #10 in News Commentary on Apple Podcasts. before this, I spent years turning other companies' conversion rates into revenue, $150 million in added revenue to be exact, wrote a book about it, and got a wireframe method design-registered in Norway twice. now I do it for my own company. bootstrapped the journey so far.

A clever first line doesn't hide a send pattern from the inbox. I'd bet on outbound where finding the next right person takes more work than adding more names.

A progress bar says the system still has work to do. If the draft is waiting for review, that screen is telling the wrong story. Mato’s Sept. 14 changelog describes manual-mode generation recognizing review as a completed step and taking the person to the episode. Show the review state where the person can act. Don’t leave them guessing whether generation stalled.

Every agent benchmark has one thing in common. Somebody already knows the right answer, which is what makes it scoreable. Most of the work people actually want to hand off doesn't come with an answer key. Pricing a new plan. Deciding which customer complaint is the one that matters. Judging that an episode isn't good enough to publish. So a model topping a leaderboard is telling you it got better at the checkable part of your job. That was never the part eating your week.

Color on a product page usually gets spent on the CTA. That’s how so many pages end up shouting the same thing. In Mato’s Sept. 1 visual-system post, the gradient waits for “interview,” “finished episode,” and “out loud.” Warm white and graphite do the quieter work around it. I’d rather use the loudest color to mark the moment a person matters than to point at a button.

Pronunciation can pass while a conversation falls flat. Every show we’d produced so far went through third-party TTS. The voices were clean and the words were right, but our team kept hearing two hosts read in turns. That was the problem Mato Voice was built to address. Score accuracy, then listen for whether the second host sounds like they heard the first. A transcript alone can’t grade that.

@speedrun ✅

third time's the charm @ycombinator? #YCW27

A podcast budget gets misleading when it spreads the launch bill across every episode. Equipment, artwork, feed migration and initial templates may matter at launch. They’re separate from the recurring work to publish episode 24. Price the production hours at a loaded rate, add recurring outside costs, then multiply by your publishing cadence. I’d keep launch spend out of the per-episode number. Averaging it in hides what the next episode actually costs.

Readers shouldn’t have to reverse-engineer a launch screenshot. Our Oct. 5 Mato Recordings post showed real Mato UI. The 45-minute IBM Research keynote in the example was public material from Wikimedia Commons, used as sample data. The screenshot showed the Recordings hub after it planned 12 things from that file. That label lets people judge what the product did without guessing whether the example came from a customer. Put the sample-data note beside the asset, not in fine print.

A voice demo needs more than three clips to play. You need to know what stayed the same. For our published Mato Voice comparison, we rendered one two-host thermostat script through Gemini, ElevenLabs and Mato Voice, with identical speaker turns, matched loudness and no timeline edits. If you’ve run a voice comparison, what did you hold constant?

An assistant integration earns a second use when the first question doesn’t require an ID hunt. Mato MCP can use the podcasts, talent and sources in your workspace. The example prompt is concrete: “In our brand workspace, compare the last three episodes over the past 30 days. Show the available download figures and suggest a follow-up topic. Ask me before creating anything.” If someone has to fetch internal identifiers before asking that, the integration has moved the setup work onto them.

Match loudness before deciding which AI voice sounds better. Mato Voice’s comparison runs one two-host thermostat exchange through Gemini, ElevenLabs and Mato Voice. The script and speaker turns stay the same, the clips share a loudness target, and there are no timeline edits. Change the mix and you’ve changed the test.

Before you judge an AI integration by how fast the demo connects, check what setup asks you to paste. Mato MCP connects through a hosted endpoint: sign in to Mato and approve access. The assistant needs remote MCP over Streamable HTTP with OAuth. No Mato API key to paste into chat and no local server to run. A key pasted into chat is part of that conversation. Credential handling belongs in the demo, not a footnote.

Retry determinism starts with remembering what the first attempt chose. Mato's Sept. 21 changelog describes freezing an episode's voice direction and reference, then reusing that snapshot across audio parts and retries. A follow-up fix kept it stable when job data was reordered between workflow steps, preventing affected MOSS episodes from failing before the first audio part. The retry inherits the decision. It doesn't get to make it again.

An assistant shouldn't get permission to create in a work account just because it can read from it. I'd start with read-only, then make creation a separate choice. With Mato MCP, you choose which workspaces the assistant can access and can leave creation unchecked at consent. Generated episodes stay in Mato for review before publishing. If read access isn't enough to prove the assistant is useful, write access won't fix that.

A transcript gets tested when someone asks, “who said that?” Mato Recordings labels speakers, lets you rename them, and plays the recording from any line. You can check the quote against the audio. If verification means scrubbing through the recording, the transcript still has work to do.

Salary-only math gives an in-house podcast a discount before you compare it with a vendor. Use the loaded internal rate your finance team uses, or document your assumption. The U.S. Bureau of Labor Statistics reports wages and employer benefit costs separately, so salary alone misses part of the labor cost. Multiply that rate by the hours per episode and your publishing cadence. Otherwise, you’re comparing a vendor quote with only one slice of your team’s cost.

A clip can have the best line in a talk and still feel unusable if the cut grabs half a word at either end. Mato Recordings plans clip moments from the word-level transcript. That timing matters because the boundary is part of the edit, not cleanup after the good bit is found. Finding the moment gets you halfway there. The cut decides whether anyone shares it.

Ask an expert what they do when the playbook runs out. A summarizer can't find an answer that never made it into a document. Mato's vision is an interview engine: agents prepare, listen closely, and stay with an unfinished thought long enough for it to become useful. AI knowledge capture has more room to grow in that follow-up than in another pass over the same files.