AI systems & automation
Fewer spreadsheets.Less rework.More sales.
I'm Leonardo. I build apps, integrations and AI automations that get your operation off paper, spreadsheets and WhatsApp threads. Before I write any code, we spell out together how the system should behave, using your own business data. You sign off, and every scenario becomes a test.
- 01 · INClient emailReact
- 02 · AIAI extracts the itemsLLMs
- 03 · MATCHMatch among 4,680 productsSemantic search
- 04 · AGENTAgent finds the costAgents
- 05 · OUTProposal in the ERPERP
→ email received from buyer@client
- Python
- FastAPI
- Node.js
- React
- TypeScript
- PWA
- PostgreSQL
- Supabase
- LLMs
- AI agents
- Semantic search
- n8n
- GitHub Actions
- ERP
- Pix
- SEFAZ
- Carrier APIs
- REST APIs
Problems I solve
Where your business quietly loses money
Sales and operations problems that show up in companies of every size, and what a well-built system changes in each one.
Marketplace fees eating your margin
Every order that comes in through a delivery app leaves a big slice with the platform.
Your own channel: a store on the customer's phone, with Pix payments and the delivery fee calculated for you. No commission.
- PWA
- Pix
- Python
- React
Slow manual quotes that lose the sale
A client asks for a price by email and the answer goes out the next day, after they've already bought elsewhere.
Quotes in minutes: AI reads the request, finds the products and prices, and the proposal comes out ready for your ERP.
- Applied AI
- Semantic matching
- ERP
- FastAPI
Leads going cold for lack of follow-up
The contact list exists, but nobody has time to write, chase replies and remember the next step.
Prospecting agent: it sends the sequence, reads the replies and hands your sales team only the people who are interested.
- AI agent
- n8n
- CRM
- PostgreSQL
Paper-based operations and shipping errors
Goods leave unchecked, with the wrong measurements, and nobody checks the freight you get billed.
Operator app: receiving, picking and shipping from a phone, with freight quotes checked against the CT-e.
- Operations app
- ERP
- Carrier API
- TypeScript
Spreadsheets that don't talk to your ERP
The same data gets typed in two or three times, and every copy has a different mistake.
Integration: your systems exchange data on their own, with a log of everything that went through.
Waiting on someone for every small change
A minor tweak turns into weeks of waiting and a new quote.
Agent-handled tickets: you ask from inside the system, and the fix comes back tested, with the previous version kept.
Work
Every system is a flow
Four systems in use, built for real problems. Follow the data through each one, step by step. Client names, companies and figures in the screenshots and demos have been replaced with fictitious data.
Quoting desk · B2B distributorIn use
From a client's email to a finished proposal in the ERP
The quoting team drops in a client's request and AI builds the proposal: items identified, products matched in the catalog, costs looked up, and a file ready for the ERP.
- 01 · INClient emailReact
- 02 · AIAI extracts the itemsLLMs
- 03 · MATCHMatch among 4,680 productsSemantic search
- 04 · AGENTAgent finds the costAgents
- 05 · OUTProposal in the ERPERP
- 224
- proposals last month
- 65%
- of price suggestions accepted without correction
- 4,680
- products in the catalog
Ordering app · bakeryIn use
A direct ordering channel with Pix, and no marketplace commission
The bakery's own store on the customer's phone, with delivery priced by real distance and payment by Pix. The owner runs everything from one dashboard and asks for changes by ticket.
- 01 · INCustomer orders on their phonePWA
- 02 · FEEFee by real distancePython
- 03 · PIXPays by PixPix
- 04 · PANELOwner's dashboardReact · PostgreSQL
- 05 · AGENTAgent ships the fixesAgents
- 60+
- versions in 3 days
- 0%
- marketplace commission
- 2
- stock locations
Prospecting CRM · sales teamIn use
An AI agent that works the leads on its own
The agent sends the sequence, reads the replies and sets the next step. Sales reps only get the people who are interested.
- 01 · INLead listPostgreSQL
- 02 · AGENTAgent sends the sequenceAgents
- 03 · READReads the repliesLLMs
- 04 · ROUTINESchedules the next stepn8n
- 05 · OUTWarm lead to a sales repReact
- 347
- companies reached last month
- 16%
- reply rate
Logistics · distribution centerIn use
Receiving, picking and shipping from the operator's phone
Operators receive, pick and ship from their phones, with every invoice checked. Managers follow each order and check the freight against the CT-e.
- 01 · INScan the DANFESEFAZ
- 02 · AIAI matches the itemsLLMs
- 03 · PICKPick with measures & photoTypeScript
- 04 · SHIPScan the DACTEERP
- 05 · FREIGHTFreight vs CT-eCarrier APIs
- 6
- pipeline stages tracked per order
- 3
- fiscal documents checked (DANFE, NF-e, DACTE)
- Input
- the quoting team drops in the email, spreadsheet or screenshot of the order.
- AI
- identifies each item and finds the right product among 4,680 in the catalog.
- Pricing
- an agent looks up supplier and cost; 65% of its suggestions go through without correction.
- Result
- 224 proposals last month, exported in the ERP's own format.
- Customer
- orders on their phone, enters a postal code, sees the fee for the real distance and pays by Pix.
- Owner
- a dashboard for orders, production, stock across two locations and scheduled orders.
- Maintenance
- the owner opens a ticket, and an AI agent fixes, tests and ships the change.
- Result
- 60+ versions in 3 days, all from the owners' requests.
- Sequence
- the agent sends the emails, reads the replies and sets the next step.
- Pipeline
- a dashboard with a kanban board, accounts per rep and the agent's activity.
- Routines
- automated tasks in n8n; the agent carries them out, postpones them or declines with a reason.
- Result
- 347 companies reached last month, with a 16% reply rate.
- Operator
- receives by scanning the DANFE, picks with dimensions, weight and photo, ships with the DACTE.
- AI
- matches what arrived at the warehouse with the item on the sales order.
- Management
- a dashboard per order, with carrier freight quotes checked against the CT-e.
- Result
- nothing ships without a checked invoice and recorded measurements.
Try it
Try the systems yourself
Short, clickable samples of a few projects, running on fictitious data. Click, fill things in and see what happens. Nothing here is sent anywhere.
Quoting desk
- Click Process email and watch the AI read the order
- Change the markup and watch the total update
- Export to see the file that goes to the ERP
Items, suppliers and prices are fictitious.
Ordering app
- Add two or three flans
- At checkout, use the sample postal code and calculate delivery
- Generate the Pix code and simulate the payment
Sample prices and address. The Pix code is fake and charges nothing.
Agent-run CRM
- Tap a lead to see the agent's history
- Use the agent's Next step and watch the lead move down the pipeline
Companies and conversations are fictitious.
Distribution center
- On the phone, scan the invoice and check the items
- Confirm the AI's suggestion, pick, and record the package
- Scan the DACTE and watch the manager's dashboard update
Invoices, orders and customers are demo data.
Method
You know what you're getting before you pay for the build
The same process on every project, from the first message to maintenance.
Written interview
Straight questions, answered in your own time, no long meetings. You tell me how the business actually runs today.
Approved scenarios
I write each behavior as a scenario, using your real data. You approve scenarios, not technical documents.
Built with tests
Every scenario becomes an automated test. Nothing goes live until it passes all of them.
Documented handover
Video manual, code and accounts in your name from day one. No hidden monthly fees.
Changes by ticket
You request changes through a ticket. The agent prepares and tests them, and keeps the previous version in case you need to roll back.
What a scenario looks like
This one comes from the bakery app, approved by the owner before anything was built. It becomes the test that keeps the rule true for good.
- Given an order for delivery 4.2 km from the store
- And the 3 to 5 km band costs R$ 14.00
- When the customer enters their postal code and calculates delivery
- Then the summary shows "Delivery R$ 14.00"
- And the total adds up the items and delivery

About
I'm Leonardo Barrey, a developer based in Florianópolis, Brazil. I build AI systems and automations for businesses, and I write the scenarios with you before I write any code.
- Let's build it
- From paper to system
Contact
Got a process that still runs on spreadsheets or WhatsApp?
Tell me how it works today. I'll come back with what can be automated, what it costs and how long it takes.
+55 48 99940-6747