Blog

Nine months of Call Copilot in one breakdown: the full funnel from install to payment, €517 of ads that brought zero sales, and why the most active customer loses me money.

Overnight I lost a user who had made it through the whole funnel: Deepgram wouldn't let him record, and the extension showed a raw technical error. Working through the case, I realized the analyst agent that explained it to me could be a product in its own right.

A week of hunting for external traffic: a directory where I'm 144th in the list of alternatives, and Reddit, where the first post is a month of karma-building away. The conclusion — the only channel that depends on me alone is my own hub.

A lead from the US got further down the funnel than anyone before him — and stumbled over microphone access. Two and a half hours on one email, written by hand — and the first hint of a profile of my target audience.

Call Copilot has a local mode: recognition runs right in the browser and the audio never leaves the device. A handful of people switch it on, nearly all of them in the EU. The privacy bet hasn't paid off.

Shipped 1.9.10 with a fully simulated demo and fixed onboarding — and the stats flipped negative: almost 30 people gone in 3 days. Instead of chasing growth, I cut about 700 euros a year in costs.

How I moved from prompting task after task to a cycle of plan → tasks → review agents → validation. Where this construction wins, what the review phase costs, and why throwing E2E tests out of the cycle was a mistake.

I gave an agent access to my Hetzner account through the browser, to the server over SSH, and to GitHub. It walked through the whole infrastructure, broke nothing, and wrote a report with a cleanup plan — minus €700 a year in expenses.

Hooked up analytics and watched what users do with the extension. They press the record button five times, close the sidebar, uninstall. I'm redoing the onboarding and the demo.

Shipped a version with automated emails and a new listing, then dug in over the weekend and found it: since March, Stripe had been sending the renewal webhook differently, and people's credits weren't being credited. Sat up until three in the morning fixing it.

An AI review showed that almost nothing in the extension actually works. I fixed it, shipped two versions with mandatory registration — and deliberately stopped, to let the numbers accumulate.

My neighbor doesn't use AI and didn't want to look at anything — then started trying it himself. On the grocery-list demo, pricing, presets, the App Store and one nasty microphone bug.

The week's main result: the Ukrainian association's site is public. SEO, sharing, sending mail and documents straight from the site — all working. The contract said end of June; we opened in mid-July.

Stripe is wired up, a test payment went through on production, self-hosted analytics is running on my own server. A week of work on the hub — and from the outside the site looks exactly the same as before.

Anthropic's new, strongest model — Fable 5 — became available for a week — and the first thing I did wasn't write code with it. I pointed it at reviewing my changes. It found a staggering number of bugs and drew up a fix plan itself.

I sat down to look at a dormant product: Supabase at 35 dollars a month, 75% of tokens going to free traffic, zero payback — and a hundred new users a month from organic growth. Deciding what to do with it.

I'm hauling dictation into the hub and trying to join three products on three different technologies into one subscription. So far the one thing their economics have in common is that it's negative everywhere.

My test client remembered the subtitle editor by herself and made subtitles for her video by herself. Next up: drop the shared password, wire up sign-in and payments.

I finally went through Call Copilot's install statistics and found out that three quarters of the tokens are burned by free and anonymous users. I switch off the free plan, introduce registration and a hard paywall — and can't properly test the thing I just built.

The Ukrainian Verein's site reached its last phase: I added a simple CMS, so they now manage almost all the content themselves. What's left is waiting on their texts and photos.

A friend commissioned a dead-simple dictation app for his elderly client. My shortcut runs into the background-recording wall — and a prototype of the new app came together in seventeen minutes.

The second saltwater fishing trip: fewer fish, more sun, two sleepless nights on the road — and the same effect that made it worth going.

In one month transcription cost about 180 dollars, and free users burned three quarters of that. Meanwhile the fourth purchase in the project came in at 2 euros. I decide to shut the free plan down.

I spent an hour writing a plan, the agent worked for four hours — by evening the subtitle editor had a mobile version: video from the iPhone photo library, subtitles, burn-in and sharing the finished clip.

Added from-scratch subtitle generation to the editor: transcription, cleanup, translation — all right in the browser. And the home task the project started with is finally closed.

Four hours without a break — and the subtitle editor is deployed to a domain that had been dead for years. The free half works, the paid half is planned, and $250 is set aside for ads.

An ordinary solo-dev weekend with the family: pet shops and ice cream with the kids, an airport run at five in the morning — and agents writing code while I push the swings.

Six weeks with almost no movement: updated the paywall, signed up for an accelerator, and then the extension lived on its own for two weeks and suddenly brought in 24 euros. The third purchase — 1.99.

I gave an agent a browser and asked it to spend my Audible credits on books it had picked itself. It bought the first one with a credit, the rest — with my money: almost €200. A breakdown of where exactly I went wrong.

I went through a year and a half of my voice notes and counted sixteen projects: some in repositories, some on subdomains, some living only in my memory. I decide to build one hub: a single domain, products in categories, a blog as the foundation.

Back in September someone passed my name to a Ukrainian association in Germany. Today was the first real meeting: requirements, fonts, a domain — and I prepared for it with Claude.

Tested an AI-dictation idea: built an iPhone shortcut and gave it to my parents — they use it, though people close to you don’t count. Set up the same dictation on my MacBook: I speak in any language and get clean text in English.

Ads are off, traffic is organic only — and on March 3, 2026 the first payment from a stranger lands in Stripe. Exactly what I'd been waiting for since December. And exactly one payment.

I finally built myself proper analytics and ran the funnel math: 10% activation, almost everyone sees the paywall, zero payments. The arithmetic says turn off the ads — so I'm turning them off.

Launched ads across ten Western European countries, and people uninstall the extension saying it's not what they expected. I dig into the numbers, move from ads to SEO, and find a competitor with three thousand users.

Found out that the extension's identifier is the same across all installs, not a new one each time — I went looking for workarounds for nothing. Rebuilt attribution and went to my first tracking session.

Launched Google Ads, set up a CRM, wired up emails. A decent number of installs, a euro apiece. And then the drop-off: people install the extension and do nothing.

The extension officially hit the Chrome Store, and the first payments came from friends. And it turned out one of my servers had been mining someone's crypto for a week — courtesy of a fresh hole in React Server Components.

A Friday in the middle of a launch: short on sleep, simple tasks stretched across the whole day, everything slightly off. What I do when I feel I'm getting close to the edge.

Google's approval came in overnight, and the excitement was real — I'd been working toward it for a long time. Then: onboarding, lead forms, a stubborn Stripe, and turning down a fifty-fifty partnership.

Feedback from the launch course: the product works, but I'm not doing the one thing that matters — no label, no shelf, no ads. Figuring out how to fix that.

I bought Mikhail Tabunov's course on launching pet projects. The main idea of the first modules is surprisingly simple: don't build anything big — take an unfinished project you already have and get it to revenue.

The course is generating more ideas than I can work through — and it's déjà vu: this is exactly what I built Referity for. The product didn't take off, but the habit of talking into the journal stayed.

The Referity race is logged in the voice journal day by day, but the final isn't: fifth place, a jury question we couldn't answer, and a debrief with the organizer. Wrapping it up — and why the next run will be different.

Hung the pull-up bar, finally sorted the car for €916, cleaned out the closets, said goodbye to the €500-a-month coach — and for the first time in a long while I feel energy coming back.

Seventeen minutes into the voice journal about being torn between projects, doomscrolling Instagram, and hunting for a magic solution that was with me the whole time.

The last accelerator session before the final landed on a day when the whole team was away. We canceled even the online option — and then five days of fishing in the fjords did what no weekend ever had.

The structure frontend built in a day, a plan for a rough MVP by the end of the week — and a paid Max subscription to Claude as an investment in development pace.

The deploy is on fire two hours before the presentation, the voice agent lands fifteen minutes before it — and in the evening, out of nineteen teams they name ten, and one of them is ours. All of it on my birthday.

A Saturday with an early start, an optimist friend, and a pond where nothing bites. A day without a single line of code — one that stays in memory better than many working days.

Ran local models on my Mac: embeddings as good as the cloud ones, Mistral extracts tasks from notes. Privacy is turning from a pretty word in the pitch into a feature we can actually build.

In the morning — a new name and the Referity.ai domain with a redirect from the old Speech Journal; in the evening — the first offline presentation at the accelerator. A B grade: one thought slipped my mind and the ending got crumpled.

An email from Startbahn27: we're in the accelerator. Ahead — a kick-off and the first presentation, and that same evening a friend asked to join the team — and I wasn't sure at first whether to be happy about it.

After submitting the application I couldn't stop: three more hours — and the app became an installable PWA, in the afternoon my wife saw it, in the evening — friends. Plus the first survival tests for the recording.

A two-day sprint: the bot turned into a standalone web app with a demo mode — while submitting the accelerator application itself took fifteen minutes and ended with a beer.

Invited a friend into the project — and he caught fire even more than me. We sat down to prepare the accelerator application and recorded the whole brainstorm straight into the journal: target audience, integrations, and the biggest problem — the input interface.

Came to a coworking space to get some work done — walked into a student project pitch session, met a professor, pitched my voice journal in German for the first time, and left with the contacts of a local accelerator.

The voice journal is three weeks old: the habit of narrating my day has stuck, my friends got a ten-minute video, and for the first time I'm recording my year-in-review by voice — into the journal itself.

A holiday week in stolen moments: semantic search via embeddings, a date filter, a chat that links back to the source notes. How the voice journal works inside — and why it's time to stop improving it.

Two sleepless nights — and my Telegram bot now takes voice notes, stores the files, and transcribes them into text. Why I need this at all — I recorded straight into it.