moul4z

Daniel Moulaz @moul4z

I build automations and AI agents that run in production at MBC, an online course company in Brazil. There, I run the assistant that checks the math before answering labor-law calculation students, and I built the system where every sale becomes an enrollment without anyone typing a thing. At home, I build Nexo, the app I manage my own money with.

commits
1,144
days with commits
102
on the busiest day (Sep 26)
71
days in a row
12

fig. 0 · my commits per day since June 2026, in MBC's repositories and my own. Hover or tap a block.

Based in
Minas Gerais, Brazil
Now
MBC Negócios Digitais, remoteInformation Systems, UFOP
Looking for
Internship or junior role
Contact
danielmoulaz7@gmail.comGitHubLinkedInResume

01 / 05

Projects

My own projects, with the code open on GitHub.

In use

Nexo

A personal finance app for Android and Windows that I use every day. I started it on 10/28/2026 and built it alone: 299 commits up to version 1.17.

Flutter · Dart · SQLite · Supabase · MCP

91 test files. The server is tested offline, with fictitious data only.

  • Data stays on the device (SQLite) and syncs between phone and PC through Supabase.
  • Open Finance: whatever shows up at the bank becomes a suggestion to review. Nothing becomes a transaction on its own.
  • An MCP connector with 98 tools so Claude can do in the app what I do on screen. Every request can be undone, and a test fails if the app gains an action without a tool.
  • Real bank rules: a purchase on the closing day goes to the next statement, and paying the statement doesn't count as spending twice.

Live

SEPTEM Cinemas

From seat to entrance: showtimes, concurrent seat booking, QR code tickets and a simulated payment.

React · TypeScript · Fastify · Prisma · PostgreSQL

Concurrency control in PostgreSQL and 300+ automated tests in CI.

Also built

  • Aresta

    Code ↗

    A webhook and integration dashboard: events, idempotency, retries and failure diagnosis.

    HMAC signatures, idempotency, retries and a dead-letter queue.

    Python · FastAPI · PostgreSQL · Docker

  • SupportFlow AI

    Code ↗

    An API that triages support tickets: category, priority, summary and a suggested reply.

    OpenAPI contract, Zod validation, tests and CI.

    TypeScript · Fastify · Zod · OpenAPI

02 / 05

Work at MBC

Since 06/2026. The code is private, so I explain what I did without exposing clients or students.

  1. 71pull requests

    AI assistant for labor-law calculations

    Answers students' questions about PJeCalc, reads the court decision as a PDF and checks the math before answering.

    React · TypeScript · Edge Functions · PostgreSQL · OpenAI

    Read the case
    The problem
    Students in a labor-law calculation course get stuck on PJeCalc (the courts' calculation software) and on parts of the decision, often outside support hours. Here, a wrong answer becomes a wrong calculation in a real case.
    What I did

    I've owned the project since July 2026: 300+ commits and 71 pull requests, up to version 9.5.50.

    • Migrated the Python/FastAPI prototype to React, TypeScript Edge Functions and PostgreSQL.
    • Two-step answers: a draft, then a reviewer that checks the legal grounding and returns a structured approval.
    • Reads PDFs and screenshots of court decisions, with a dedicated mode for analyzing the decision.
    • Automatic checks on the math that appears in the answer.
    • Students only: sign-in by email link or code, checked against a student list synced from the membership platform.
    • Regression tests on every version and 60+ reviews of real conversations.
    The hard part
    Cost without losing quality. I measured that only 14% of the prompt hit the cache, because the fixed part was mixed with each question's context. I moved the fixed part first, and the reviewer now just approves a correct draft instead of rewriting the whole answer.
    Result
    In production for students, with the cost of each answer measured (about US$ 0.04 in the August review) and every fix tracked in a version.
  2. 87commits of mine

    Base MBC, the company's internal system

    Experts, products, offers and enrollments in one place, with sales coming in on their own.

    React · TypeScript · Supabase · PostgreSQL · Edge Functions

    Read the case
    The problem
    Sales, access and support data for each business unit lived on different platforms, and every report was put together by hand.
    What I did

    I built most of the system between August and October 2026.

    • Guru and TMB sales arrive by webhook and become enrollments through functions in the database itself.
    • Access sync with the membership platform (Curseduca).
    • A support dashboard with Chatwoot data: satisfaction (CSAT) and time to first response, recalculated every day.
    • A role-based permission matrix covered by tests, with personal data trimmed according to the viewer's role.
    The hard part
    Security. The browser can't read any table: every table has RLS with zero policies, and all access goes through an API that checks the login, team membership and whether the role allows the action. Revoking someone's access takes effect on the very next request.
    Result
    In production. Before every release, the check runs types, lint, tests, build and a secret scan.
  3. 1,620checks

    A prospecting CRM built on court gazettes

    Turns labor court publications into leads, with pricing, proposal and contract in the same place.

    Python · JavaScript · PostgreSQL

    Read the case
    The problem
    Students prospect law firms, but the information is spread across the daily publications of 24 labor courts, and everything else lived in spreadsheets and separate tools.
    What I did

    I built most of it (101 of 150 commits), together with one more person on the team.

    • Daily collection of publications from 24 labor courts, with a dedicated PostgreSQL index so search is interactive.
    • Leads linked to people and firms by bar number (OAB) and company ID (CNPJ), with a Kanban board and Google Maps as a second source.
    • Pricing, proposal and contract built inside each lead.
    • Sign-in by email link.
    The hard part
    The public API only answers per court, per day and per phrase, capped at 1,000 items. A live search would take dozens of calls, so I split batch collection from a dedicated index where search is interactive.
    Result
    In development. Every phase ships with its own test suite, now at 1,620 checks, plus mutation testing on the critical rules.
  4. 7units (4 today)

    AI support agents

    Agents on WhatsApp and Chatwoot that know when to hand off to a human.

    n8n · Claude · OpenAI · Chatwoot · PostgreSQL

    Read the case
    The problem
    Human support answered the same questions every day, in every MBC business unit, on WhatsApp and Chatwoot.
    What I did

    I built and maintain the agents, the knowledge base and the curation flow.

    • n8n agents using Claude and OpenAI, connected to WhatsApp and Chatwoot.
    • A knowledge base built from real support answers, not generic docs.
    • Handoff to a human agent without duplicate replies.
    • A curation panel (Vercel + Postgres): a workflow finds new questions in conversations, and the team approves them before they reach the base.
    The hard part
    The handoff. When a person takes over, the agent has to stop right away, without sending anything else and without leaving the customer unanswered.
    Result
    I built the agents for 7 business units. Some later left MBC, and 4 are active today, with human handoff. Nothing new reaches the base without the team's approval.

Also at MBC

  • Integrations for 5 payment platforms (Guru, TMB, Hubla, PagTrust and Kiwify) with the membership platform and WhatsApp. If reading a webhook fails, nobody loses access.
  • Full support setup for new business units: Chatwoot, WhatsApp, tags, tickets, an AI agent and a cancellation playbook.
  • Launch automation: lead-capture groups, reminders through the official WhatsApp API, cart-open messages and product release on purchase.
  • Internal tools: a bonus release tool used by 4 business units, a class scheduler and disconnect alerts for 5 WhatsApp inboxes.

03 / 05

Tools

What I build with, linked to the projects on this page.

Pick a tool to see which projects use it.

Languages & UI

Server & data

AI & integrations

Quality

04 / 05

Experience

Work

  1. 06.2026 – now

    Operations, Automation & AI Intern

    MBC Negócios Digitais · remote

    Applied AI, internal systems and integrations. I work with 4 business units today.

  2. 10.2026

    AI Evaluation Specialist (Portuguese)

    Handshake AI · freelance

    Evaluated model responses against English guidelines. Every submission approved.

  3. 09.2025 – 03.2026

    Sales Apprentice

    Pacaembu Autopeças

    Customer service, targets and Excel reporting.

  4. 11.2023 – 08.2024

    Administrative Apprentice

    Grupo Friopeças

    Administrative and finance processes in SAP.

Education

  1. 03.2026 – now

    B.S. in Information Systems

    UFOP

    Second term: C# and object-oriented programming, data structures in C, and information management.

  2. 2026

    Data Analysis & AI

    Santander + Alura/FIAP

    120h foundation track completed and 148h Boost track, 17 of 18 courses (SQL, Python, Power BI).

  3. 08.2024 – 08.2025

    Industrial Apprenticeship in Industrial Management

    SENAI · 911 hours

Languages

Portuguese (native) · English B2 · Spanish (basic)

05 / 05

Contact

Email is the fastest way to reach me.