What we built for ourselves

Five of our own tools with traffic measured across three thirty-day windows, including one where growth did not happen and the release was rolled back. With the counting method and the caveat about seasonality.

Five of our own tools. Three carry a link — those are clean, broad-audience tools you can open and check. Two deliberately do not: one is being reworked, the other we do not put on a public showcase. The numbers are visits across three thirty-day windows; the method is below.

Translating text from a photograph

49,220 → 140,444 visits per month +185%

The problem.: Someone photographs a sign, a price tag or a page and gets a translation. The hard part is not translation — it is recognition under bad conditions: handheld, glare, at an angle.
What we did.: The image is downscaled in the browser before upload; recognition and translation run as two task profiles instead of one general-purpose model; a repeat request for the same file is served from cache.
How it ended.: Our largest tool by traffic after the text one. Its weak spot is stated plainly in our own write-up: pages per visit around 1.3 — arrive, press, leave.

Open the tool

Weather with nothing extra

3,158 → 19,133 visits per month +506%

The problem.: A crowded niche with strong incumbents. You cannot win it on forecast quality — everyone draws on the same source. You win on speed and on how much clutter a person sees before the answer.
What we did.: Extra blocks removed on purpose; the answer is rendered server-side, ready. This tool makes no model calls at all — verified in the code, not claimed. It is our answer to the question «does this actually need a model».
How it ended.: Sixfold growth in three months. We keep it on this page precisely because it is a case about declining to use AI where AI is not needed.

Open the tool

Translating audio and speech

810 → 11,528 visits per month +1323%

The problem.: Audio or video in, text and translation out. The difficulty is not the model but the length: a ten-minute file does not fit in one request and must not hold a connection open.
What we did.: Asynchronous handling: the job is submitted, the page polls for the result, and there is a marker for «a human has seen the answer». A long file no longer looks like a hung page.
How it ended.: Fourteenfold growth from a low base. This is the most honest shape of growth to discuss: it shows not the power of the tool but that the niche was found before the competition got there.

Open the tool

Scoring a photograph

82,541 → 165,306 visits per month +100%

The problem.: A mechanic where the user gets a score and a comparison. Entirely about engagement: a single answer is useless here, what matters is the second and third attempt.
What we did.: One thing was built that neighbours in the niche do not have — a comparative mechanic instead of a single verdict. Everything else was deliberately left alone.
How it ended.: Doubled in three months, and traffic is not the important number here: 4.2 pages per visit and nearly five and a half minutes per session, against 1.3 and about a minute on the neighbouring tools.

A large text tool

955,533 → 828,186 visits per month -13%

The problem.: Our most visited tool. We shipped a new version of the text processing expecting a gain in quality.
What we did.: We collected the feedback for the period and READ it, instead of looking only at traffic. The complaints were about the QUALITY of the answer — not speed, not appearance. The release was rolled back to the previous version and the new one sent back behind a flag that enables it for a fraction of traffic.
How it ended.: Down 13 % over three months — the only negative case on this page, and it is left here on purpose. What we took from it: changing something that already works carries a higher burden of proof than building something new.

How we counted

Windows: Three thirty-day stretches: 12 May – 10 June 2026, 11 June – 10 July 2026, 11 July – 9 August 2026. The cards show the first and the last.
The metric is visits: Not rankings, not impressions, not share of search. The reason is inconvenient and therefore stated: the traffic-by-source breakdown only started being collected in our analytics in July, and for some of these sites organic traffic for May and June is recorded as zero against tens of thousands of visits. Taking «organic growth from zero» would have produced a case whose entire figure is a gap in our instrument. Visits were measured continuously across all three windows.
What we do not claim: The season. Demand in some of these niches rises into summer by itself, and this measurement cannot separate that from our work. Where growth came off a low base, it says so: fourteen times from eight hundred visits and fourteen times from eighty thousand are different statements.
What is not here: Client work: it sits under NDA, which is why there is no logo wall on this site and will not be while that holds. All five cases are our own products — our money and our risk.

If you want these same figures on screen in an analytics dashboard, with the site names attached — that is a conversation in a meeting, not a page on a website. Write to us and we will show you.

Every tool above runs through the gateway — how it works and how to get access.

Questions people ask about these

Why do only some cases carry a link?

Links go to clean, broad-audience tools: you can open them and check for yourself, which is more honest than any description. They do not go to narrow tools where reading the output takes expertise — there the cost of a wrong reading falls on you, not on us — nor to anything currently being reworked. Every link is marked as not passing weight.

Is this client work?

No. These are our own products: our money, our risk, our decision to stop. Client work sits under NDA, which is also why you will find a mechanism and a measurement here rather than a wall of logos.

Why visits, and not rankings or organic traffic?

Because visits were measured continuously across all three windows, while the source breakdown only started being collected in July — for some of these sites May and June are recorded as zero against tens of thousands of visits. Reporting growth «from zero» on that data would present a gap in our instrument as a success of the site. We have a separate note about exactly that mistake.

Why include a case where nothing grew?

Because it is the most honest one here and the most useful for a conversation about your work. A page of nothing but successes says nothing about a supplier — everyone has those on a showcase. What tells you something is what we did when we made things worse.

How does this relate to getting access to the gateway?

Directly: every tool here runs through the same gateway we offer access to. Task profiles, cached answers for repeat requests, asynchronous handling for long jobs, per-day ceilings — those are not features written for a sales page, they are what these five products needed to stay affordable to run.

Could the growth just be seasonal?

Partly, and we are not going to pretend otherwise: three windows in late spring and summer cannot separate our work from the season. What the season does not explain is the direction differing between products measured in the same months — one fell by 13 % while another grew sixfold. Treat the numbers as our own measurement, not as proof of a method.