AI product development · consulting
AI product development, all the way to real users.
I help founders and teams decide what is worth building with AI, ship the first version, and find out whether anyone wants it. Hands on the code, not on slides.
AI product development services
From diagnostic to
a running product.
Three ways to work together. If the question is whether AI solves it, start with the diagnostic. If that is settled, we go straight to building.
-
AI product diagnostic
For when you have an idea, or a request from the board, and do not yet know if AI solves it. You leave with a written decision: what to build, what not to build, what it costs to run, and how to tell if it worked.
- Use case and value hypothesis
- Prototype on your own data
- Cost per task and risks
- Scope of the first version
-
AI MVP development
End-to-end AI software development. I build the first version with you: product, models, evaluation, and launch to real users, not a demo that only works in the meeting.
- Web or mobile product, live
- Evals from day one
- Cost limits and observability
- Code and accounts in your company’s name
-
AI agent development
For a product or operation already running: support, sales, back office. Agents that answer only from what you approved, can say they do not know, and hand over to a person at the right moment.
- Approved sources and clear limits
- Handoff to a person
- WhatsApp, web, or your own system
- Weekly review of conversations
How it works
One week at a time.
30 minutes
The conversation
You tell me the problem. If I cannot help, I say so on the spot.
Week 1
The decision
We pick one use case and the number that proves whether it worked. Everything else stays out.
Weeks 2 to 5
The first version
I build the simplest version a real person can use. You see it move every week.
Weeks 5 to 8
The launch
It goes live with evals, a cost limit, and someone actually using it.
After
The handover
Code, accounts, and docs go to your team. Or we keep going, month to month.
Where AI products stall
A demo is easy.
A product is something else.
What I learned shipping my own. It applies to any AI product development lifecycle.
Starting from the model, not the problem
The first question is what the person is trying to do and what happens when the AI gets it wrong. Much of what seems to call for an agent is solved by a search box and a form.
Changing things without knowing if they improved
New model, new prompt, new knowledge base. Without a set of real cases to compare against, you only find out it got worse when a customer complains. That is why evals go in on day one.
Not measuring cost per task
The usage bill grows with success. Your price has to fit the cost of each task before you scale, not after.
Promising more than the AI knows
At Mural Vivo, the agent answers only from what the business registered. Trust is lost in one invented answer.
Thinking the code is the hard part
On tico, Apple’s review, battery life, and sync were not in the plan. And distribution is still the hardest part: AI-generated posts keep you present, but real people are what make someone stop and look.
What I have built
Real products,
stumbles included.
More than 12 years in data, product, and machine learning, including Shopee, alt.bank, and aMORA. Each product below has its story written down, including what did not work.
| Product | What it is | What it taught me |
|---|---|---|
| tico | iOS app for families with newborns, with AI | Content made by people beats what AI generates on its own. |
| Mural Vivo | AI agent that answers customers on WhatsApp from what the business taught it | Ads burn money too fast. Distribution is the real bottleneck. |
| cambio | Open-source quantitative tool for exchange-rate decisions | Taking emotion out of a decision beats trying to predict the market. |
| apibjj | Verified jiu-jitsu records, organized in a spreadsheet | Collecting and reselling data is far harder than it looks. |
| CargoFax | Customs trade data for the Americas | Software engineering to make trade data easier to use. |
Questions
What people
usually ask.
What does an AI product development consultant do?
Helps you decide where AI produces results and gets that into production. My work is product-focused: instead of a report, you leave with a written decision and, when it makes sense, with software that is live.
How much does AI product development cost?
It depends on scope, integrations, and how much risk the use case can take. I quote after the first conversation, which is free. Also budget for the recurring cost of the models, which we calculate per task during the diagnostic.
How long does it take to build an AI MVP?
With a tight scope, 4 to 8 weeks to a first version with real users. Legacy systems and app stores add time: on tico, Apple’s bureaucracy took more than three weeks per review round.
Do I need to train my own model?
Almost never at the start. Most products begin with a model behind an API, good data, and good evals. Your own model comes in when cost, privacy, or quality prove it is worth it.
AI consultant or AI development company: which do I need?
A company brings a team and a process, which suits large projects. With me, you talk to the person writing the code and making the product calls. If the scope is more than one person can handle, I tell you in the first conversation.
Is my data safe?
We choose provider, region, and retention together before any data is sent, and personal data only goes in when it is necessary. Each provider’s terms and your legal basis get reviewed with your counsel.
Who owns the code?
Your company does. Repository and accounts are in your name from day one.
Do you work with teams outside Brazil?
Yes. I work remotely, in English and Portuguese, from São Paulo.
Let’s talk
Got a product
in mind?
Three lines are enough: what the product does, who it is for, and where it is stuck. I’ll tell you honestly whether I can help, and how.
araujo.vitorgabriel@gmail.com