Mauricio Belmonte

I build C#/.NET and Python backends for ERP, tax rules and integrations.

Work

2020 to 2026
The first four have a section on this page.
ProjectWhat it isStack
SIGE Cloud ERP ERP with Brazilian tax rules, since 2020 C#, ASP.NET, Angular, MongoDB
Data migration Between ERPs, at SIGE Cloud Python, FastAPI, LLMs
BM Menu Runs a pizzeria in production .NET, Svelte, React, Postgres
Ledger Personal finance with an LLM .NET, Angular, SQLite
Local assistant Coding fully offline, in the terminal or on Telegram Python, Ollama, Qwen
COMVIDA (opens in a new tab) Landing page for a clinic in Porto Alegre Static Next.js

Ledger, personal finance.

Each expense gets its category from the cheapest layer. The LLM only sees what is left.

  1. Expenses come in from several places.

    Statements, a message or a receipt photo on Telegram. Each one becomes a transaction with no category.

    Sources: card statement, bank statement, Telegram and receipt photo.

    DateDescriptionAmount (R$)
    02/09UBER *TRIP23,90
    03/09IFOOD *RESTAURANTE45,50
    05/09PADARIA PAO QUENTE12,40
    08/09NETFLIX.COM39,90
    11/09DROGARIA SAO JOAO32,17
    A sample statement, still without categories.
  2. First, a fixed dictionary.

    UBER becomes Transport, IFOOD becomes Food. The obvious cases are settled here, with no calls at all.

    DescriptionCategory
    UBER *TRIPTransport
    IFOOD *RESTAURANTEFood
    PADARIA PAO QUENTEno rule
    NETFLIX.COMSubscriptions
    DROGARIA SAO JOAOno rule
    Fixed dictionary: UBER, IFOOD and NETFLIX.
  3. Then, what has already been learned.

    Every decided category is stored. A merchant that has shown up once is never asked about again.

    DescriptionResult
    PADARIA PAO QUENTEFood, already learned
    DROGARIA SAO JOAOnot found
    Lookup in the CategoriasAprendidas table.
  4. Only the leftovers go to the LLM.

    The LLM gets just the description and the possible categories. Its answer goes into the cache.

    DROGARIA SAO JOAO

    • Food
    • Transport
    • Groceries
    • Health
    • Leisure
    • Subscriptions
    • Shopping
    • Education
    • Services
    • Other

    Answer stored in CategoriasAprendidas.

    What the LLM receives, and what it picks.
  5. The result lives on an always-on screen.

    A kiosk dashboard at home and a bot: typing “mercado 50” on the phone creates the transaction.

    September

    • Food
    • Groceries
    • Transport
    • Health
    • Subscriptions

    TelegramTelegram

    mercado 50

    Transaction created.

    Kiosk dashboard at 1280×720 and the Telegram bot.

Built with .NET, C#, Angular and SQLite, plus a Telegram bot. Sample rows and amounts.

BM Menu runs a real pizzeria.

It started at a pizzeria where I worked as a delivery driver. Instead of replacing its system, I integrated with it: today it is two .NET services in production.

Visit bmmenu.com.br

  1. 1 Online order
  2. 2 Central queue
  3. 3 Dough and crust
  4. 4 Toppings
  5. 5 Route
  6. 6 Delivered
The path of an order, from the online menu to the customer's door.

Operations, for the pizzeria

  • Receives orders from Cardápio Web by webhook, with polling as a fallback.
  • Sorts the queue by when assembly has to start.
  • Runs assembly as a line of stages, one tablet per stage.
  • Groups deliveries into routes, with a map and a QR code for the rider.
  • Prints tickets through an agent wired to the thermal printer.

Svelte, .NET, C#, Postgres and Docker.

Platform, for the provider

  • Console for stores, licenses, plans and provisioning new stores.
  • Built to license more than one product.

React, .NET, C#, Postgres and Docker.

AI agents, the same rules.

I use agents every day to move faster. Whatever goes to production meets the same standards as always.

Agents I use: Claude Code, OpenCode, OpenClaw, ChatGPT, Gemini and Velua.

ClaudeCLAUDE.md, BM Menu operations
## Five backend invariants
1. Every write goes through a single-consumer queue.
2. The server clock is the only truth.
3. Authorization fails closed.
…
Instructions in every repository
Commands, architecture and the rules that can't be broken, read by the agent before the code.
Docs split by topic
Overview, domain, data model, operations and a page just for known gaps.
Automatic guardrails
Warnings fail the build, tests run in CI, a pre-push hook runs locally. Agent code goes through the same checks.
A contract tested on both sides
BM Menu's operations app and platform share twin tests over the same examples.
Portrait of Mauricio Belmonte
Mauricio Belmonte, in Rio Grande do Sul, Brazil.

I learned to code by understanding how things work.

I started in tech support and moved into development, learning on the job through research, experiments and more experienced developers.

I like systems I don't know yet: digging into the problem and understanding enough to reach a solid solution.

What I do

  • ERP and Brazilian tax rules, from NFe to CTe
  • Integrations between systems, by webhook and queue
  • Data migration between ERPs
  • APIs in C#/.NET and Python

Where I am and where I've been.

PeriodRoleWhereStack
2020 to now Developer SIGE Cloud ERP in C#, ASP.NET, Angular and MongoDB
2026 to now Data migration SIGE Cloud Python, FastAPI and LLMs
2026 to now Co-founder BM Menu (opens in a new tab) .NET, Svelte, React and Postgres
In progress Systems Analysis and Development Associate degree

Let's talk?

About a project, a role, or a technical idea. Email is the best channel.

[email protected]