Clientazo
ES EN
Message me

I build with AI, I don't just talk about it

This is how I built Luka Money, a fintech for small businesses, with AI, specs first and tests at every step. It's the same method I use with sales teams.

See Luka in beta ↗

Case 01 · Luka Money

A complete fintech, spec first.

Luka helps small businesses charge, pay and move their money, here and abroad. I designed the product, its brand and its code, working with AI agents.

Beta with test providers

It has no customers yet and does not handle real money. What you see here is engineering, not business results.

Luka architecture: the web app and landing call an API on Cloudflare Workers, which uses a Durable Objects ledger, a D1 database and providers behind adapters. Web appReact · Vite · TypeScript LandingHTML · CSS · JS APIHono · WorkersCloudflare LedgerDurable Objects DatabaseD1 · SQL ProvidersAdapters with mocks
Simplified diagram. Each external provider sits behind an adapter, so it is tested with mocks until the real integration arrives.
  • 19database migrations, always additiveLuka · October 26, 2026
  • 36automated test files in the APILuka · October 26, 2026
  • 7team roles, with a tested access matrixLuka · October 26, 2026
  • 13AI skills that guide the agentsLuka · October 26, 2026

Engineering

What matters when money is involved.

Technical decisions that take risk out of the product before the first customer arrives.

Swappable providers

Payments and identity verification come in through ports and adapters. Today they run on mocks, and switching to the real provider does not touch the rest of the code.

A ledger that stays balanced

Each merchant has its own ledger in a Durable Object: atomic balances and limits, money as integers and a mandatory idempotency key on every write.

Business rules as data

The policy (levels, limits, pricing) is generated from the financial model in a spreadsheet, and a test fails if the code and the model drift apart.

Tests that attack

The same scenarios run against two database implementations, and a second merchant tries to read every route of the first to prove there are no leaks.

Security from day one

Second factor for reviewers and withdrawals, signed webhooks with a time window, attempt limits and seven roles with verified permissions.

Monitored operation

A scheduled job reconciles and retries, and a synthetic merchant charges, pays and reconciles every hour in beta to flag anything that breaks.

Method

How to build with AI without losing control.

  1. 1

    Spec first

    Nothing gets built without an approved specification. A project instructions file keeps the rules, the decisions and what must not be broken.

  2. 2

    Agents with skills

    Agents work with skills for interface, Cloudflare, performance and security review, and every task closes with evidence, not a «done».

  3. 3

    Agents from the terminal

    I build scheduled agents with skills, MCP, SDKs and APIs. Every project starts with SDD and with orchestrated agents and subagents, each with its own role.

  4. 4

    Tests and human review

    Type checks, formatting and tests before anything is called finished, and I decide every change that touches money.

Being honest

What Luka is not yet.

It is not in production and has no customers. The real integration with the payments partner, the exchange declaration, electronic invoicing and the legal opinion are still missing. I say it this way because it is what I would do with your team: tell you what is done and what is not.

Shall we talk?

Tell me how your team sells today. We chat by message or on a short call and, if it's a fit, we follow up.