Dinimiciuil Labs runs five Ai products in parallel: Snapeto (live), Loometo (working studio), and three deliberate scaffolds waiting their turn. This post is the builder's ledger: what the shipped ones actually cost to run, the features we killed, the money we burned, and the one failure pattern we hit three separate times before naming it.
Most 'lessons from building AI products' posts are written from a distance. This one comes from our own session logs, the same way we wrote the Snapeto teardown. Every number below is a real figure from our books, not an industry estimate. Where we only have a vague memory, we say so.
How much does it actually cost to run an Ai product?
Less than most founders fear, if you route models by task. Snapeto's fridge scan costs about $0.0013 per scan, which projects to roughly $120 a month at 1,000 users scanning three times a day. That number only stayed low because we stopped sending every request to the biggest model.
The single highest leverage change was split model routing. A cheap fast model (Gemini Flash-Lite, roughly 15 times cheaper) handles the binary question 'is this even a fridge photo?', a mid model does ingredient detection, and the expensive model only runs for paid tier recipe generation. We picked Gemini over the OpenAI equivalent for the pantry categorizer on measured grounds: about half the price of gpt-4o-mini with 250 to 400 millisecond responses instead of 400 to 800. Details of that architecture are in how we ship Ai products.
Costs also came down by deleting features. Ai generated images inside the scan flow were costing €0.10 to €0.15 per scan, a hundred times the scan itself. We cut them, and the product got better, not worse (more on that below). The founder verdict on that experiment, quoted from our logs: 'Time is money. Revert back to before this idea.'
The scan itself costs a tenth of a cent. The Ai image we bolted onto it cost fifteen cents. The bill wasn't the model. The bill was the decoration.
One embarrassing but useful incident: we hit our €5 monthly Gemini spend cap mid-session, partly because video generation experiments on the marketing side were drawing from the same account. The fix was raising it to €20 and separating app spend from marketing spend. If you run one Google Cloud project for everything, your experiments can pause your production app. We wrote about full build budgets separately in what an Ai app costs in 2026.
What does Ai get wrong that you only find by shipping?
The failures that matter never showed up in testing. They showed up when real photographs, real accents, and real dinner tables hit the product. Three examples from our logs, each of which forced a design change.
- Vision models confuse lookalikes from real angles. We deleted Snapeto's pantry camera flow after finding that, in the founder's words, 'coffee from top view looks like garam masala.' Voice input went with it: speech to text turned paneer into 'panel' and methi into 'meaty.'
- Language models write unsafe defaults with total confidence. A generated recipe said to cook mutton for 5 to 8 minutes. We now enforce hard minimum cook times per protein in a shared rules file that all six of our recipe generators must pass through.
- Ai polish can destroy trust at the exact moment you need it. Our log entry reads: 'Ai generated dish images destroy trust at the I cooked this moment.' Users want to see the real dish. We replaced generated food photography with an in-app camera at the point of cooking.
The pattern across all three: the model was never the product. The guardrails, the fallbacks, and the decision about where Ai should not appear were the product. That's the core of what we mean by Ai native design.
When should you kill a feature?
Earlier than feels comfortable. Across four months of logs, the entries we're most grateful for are deletions. The Samsung mirror app is the clearest case: we built a parallel Galaxy frame version of Snapeto, deployed it live, and mirrored every change into it for about a month. Then we measured the maintenance tax against what it earned us and killed it. One responsive web app, native later via a wrapper. The log entry ends: 'Do not re-raise.'
- Voice output went through five versions across two providers before it handled Romanized Hindi acceptably. We kept that one, because users needed it. Iteration is for features users need.
- The curved wordmark on the app icon died after six failed attempts. Instagram, Spotify and X ship icons with no wordmark at all. Deletion is for features only we wanted.
- Streak gamification was removed on sight: 'no Duolingo, we'll notify users later.' Borrowed mechanics are not product decisions.
The rule we extracted: iterate on what users need, delete what only you want, and never keep something because you already paid for it. The Samsung app was our sunk cost test, and passing it felt like a loss for about a day.
Why did we rebuild the same section six times?
Because nobody had locked a direction, and iterating with Ai is so cheap that we mistook motion for progress. This is the failure we hit three separate times across three different products before we named it. On Snapeto's marketing site, one section was rebuilt more than five times in a single session and every version got busier. The founder's final verdict, verbatim from the log: 'wholly ice slop. Leave it.' The whole folder was retired.
The Dinimiciuil site itself needed three full design resets in its first session: draft one 'looks Ai slop,' draft two 'too minimalistic like Apple,' draft three finally held. Loometo's interface went through about seven styling passes before it stopped reading as generated. Same wall, three products.
Ai driven iterative design without a strong direction locked by a human produces visually busy but unfocused output. Six rebuilds in one session never converged, because each iteration reacted to a tactical note instead of a strategic frame.
The operating rule that came out of it: when a section has been rebuilt four or more times, the direction is wrong, not the execution. Stop iterating, escalate to a human decision (in our case, a sketch drawn by hand by the founder), then execute once. The version built from the sketch shipped first try.
Was building our own Ai tooling worth it?
For us, decisively yes, and we can show the receipt. Producing Ai marketing video through subscription platforms would have meant a stack that priced out around $96 a month across four tools, with credit limits that die mid-month. We built Loometo, our own node based generation studio running on our own API keys, and produced a finished three clip, 25 second UGC style ad for $7.60 in direct model costs.
Honest caveats from the same logs. We burned about €5 in one evening of failed video generations before learning that identity drift ruins characters across clips, and another €5 discovering that location scenes want short plain text prompts while character sheets want structured ones. API calls straight from the browser hit CORS walls that forced a server proxy rewrite at 1:30 in the morning. Building your own tooling is cheaper per output and more expensive per lesson. If your volume is low, the buy versus build math can flip.
What changed about how we start new products?
The last three products (Slipeto, Beepeto, Rosmate) have shipped exactly zero lines of code, and that is the lesson working as intended. Each got one bootstrap day: positioning written down, a wedge chosen, names checked against live domain availability (all three were renamed or picked because the first choice.com was taken), domains purchased, and a memory file for decisions. Then they stopped. The robotics one, Rosmate, is the subject of a separate write-up on wiring an LLM into ROS2, which is where most of its thinking currently lives.
Early Snapeto taught us the opposite habit was expensive: building UI on placeholder assumptions means rebuilding it when the real direction arrives. Now nothing gets built speculatively. The scaffolds wait until Snapeto and Loometo earn the attention, and when their turn comes they inherit a proven stack template, this blog's distribution, and every rule in this post. Starting a product now costs us a day and two domain registrations.
How much does it cost to run an Ai feature in production?
Our real number: about $0.0013 per fridge scan on Snapeto, roughly $120 a month at 1,000 users scanning three times daily. The key is routing: a model 15 times cheaper handles validation, and the expensive model only runs where quality is paid for.
Should a small team build its own Ai tooling or subscribe?
Count your volume first. Our subscription stack for Ai video priced out at $96 a month with credit caps; one finished ad through our own pipeline cost $7.60 in model fees. Below a few outputs a month, subscriptions win. Above that, your own keys win quickly.
What is the most common mistake when iterating with Ai?
Iterating without a locked direction. We rebuilt one section six times across one session and it never converged. Our rule now: four rebuilds means the direction is wrong, not the execution. Get a human decision, then execute once.
Do Ai generated images help or hurt a product?
Depends where they sit. In our food app they actively hurt: generated dish photos broke trust at the moment users compared them to their real plate, and cost €0.10 to €0.15 per scan. We replaced them with a real camera and the product improved.
How many products can a solo founder plus Ai actually run?
In our experience: two in active build, three parked as one day scaffolds. The scaffolds cost a day of positioning work and two domains each. Trying to build all five simultaneously would have meant shipping none of them.
That's the ledger so far. Two products live and earning their lessons, three waiting with their names on the door. If you're weighing an Ai build of your own and want people who have already paid for these mistakes, see how we work or start a conversation at the bottom of our home page.
© 2026 Dinimiciuil Labs. All rights reserved. Written on the build floor in Dublin. You are welcome to quote a short excerpt with a link back; please do not republish the full article without permission.