How do engagements work?

Every project starts with a short discovery sprint. You get an architecture brief, a fixed proposal and a timeline. From there we work as a dedicated senior team or deliver a fixed scope — your call.

How fast can we start?

Discovery usually begins within two weeks of first contact. Build starts immediately after the proposal is approved.

You do. Every repository, every credential, every document — from day one. No lock-in by design.

Two ways. Internally, AI-native tooling compounds our speed without lowering the bar — every line is reviewed by senior engineers. In your product, we build AI features and agents only where they create measurable value.

Fully remote with strong overlap across US time zones. We work in English and Spanish, with US and international clients.

Scope is defined in discovery and priced as a fixed proposal or a monthly team rate. No surprise invoices — changes are agreed before they are billed.

Our own product · 2026

PicSell Pro

AI product photography for people who sell clothes

A seller photographs a garment on a hanger with their phone. The engine returns a photo good enough to sell with — same colour, same texture, same garment. That last part is the whole product, and it is why this was harder than it sounds.

USD 0.04cost per photo

4×cost reduction

7engine modules

The problem

Every AI fashion tool on the market puts the garment on an AI model. That is a different job: it serves brands with catalogues and production budgets. The person reselling forty items a month does not want a model — she wants her own garment, photographed well, in the format the channel asks for, without a return three days later because the colour was not that colour.

The decision that shaped the build

Fidelity is enforced in code, not promised in copy. `engine/prompt.ts` forbids altering colour, texture and design; `engine/fidelidad.ts` verifies automatically that none of them changed and flags the result when it cannot confirm it. No competitor publishes a fidelity check, because their proposition requires the opposite. Ours only works if the garment is untouched, so the guarantee had to be executable.

Making the unit economics work

Every generated photo costs real money, so the price of the product is a function of what the engine spends. A rework of how compositing is staged took the cost from USD 0.16 per photo to USD 0.04 — the difference between a plan that covers 93 kits a month and one that covers 375. That is not an optimisation, it is the business model.

What went wrong

The free plan's watermark read PICSELLPRO.COM. That domain does not resolve — the product runs on picsell.pro — and it was live in production with nothing catching it, because the watermark renders correctly and nobody types a domain off someone else's photo. The same string was in the web app too. It is now a single constant with a guard test, which is the only reason it can never come back.

What it taught the studio

Two things we now do on client work by default. A test is not finished until you have watched it fail on the bug it exists to catch. And a claim a product makes about itself should be verified by the product, not by the person writing the landing page.

Stack

# TypeScript
# Next.js
# Hono
# Drizzle
# PostgreSQL
# Supabase
# Gemini 2.5 Flash Image
# Paddle
# Mercado Pago
# Vercel

Visit the product Our own products