const mateus = {
role: "Senior Fullstack / Applied AI Engineer",
location: "Uberlândia, MG 🇧🇷 — remote-first, EN C1",
since: { coding: 2010, professional: 2014, techLead: 2020, appliedAI: 2023 },
founder: "Mobize Tecnologia",
stack: ["TypeScript", "Node + Fastify", "React / Next.js", "Python", "PostgreSQL + pgvector", "AWS"],
shipping: ["voice & chat agent platform", "condo management SaaS", "AI deals platform"],
belief: "Specs first. Systems thinking is why my AI work reaches production.",
} as const; |
where most of the work happens |
|
Sub-900ms end-to-end conversational loops over real telephony — turn-taking, barge-in, tool calling mid-call.
|
Semantic + BM25 retrieval with Cohere reranking, tenant-isolated, grounded answers over each customer's own knowledge base.
|
|
Traces, cost attribution, quality scoring and regression suites — so model changes are measured, not vibed.
|
Multi-agent workflows and CI pipelines where AI agents open and review merge requests.
|
🔊 How a call actually flows (click to expand the architecture)
flowchart LR
A["📞 Caller"] -->|PSTN| B["Twilio"]
B -->|audio stream| C["Deepgram STT<br/>streaming"]
C -->|partial transcripts| D{"Orchestrator<br/>Node + Fastify"}
D -->|query| E["Hybrid RAG<br/>pgvector + BM25<br/>Cohere rerank"]
D -->|tool calling| F["Elasticsearch<br/>CRM · Calendar · APIs"]
E --> D
F --> D
D -->|tokens| G["LLM<br/>+ Gemini Live"]
G -->|stream| H["ElevenLabs TTS"]
H -->|audio| B
D -.->|traces · cost · eval| I["Langfuse + OTel"]
style D fill:#0ea5e9,stroke:#0f172a,color:#fff
style E fill:#1e3a8a,stroke:#0f172a,color:#fff
style I fill:#334155,stroke:#0f172a,color:#fff
The hard part isn't the LLM call. It's the budget: every hop above has to fit inside ~900ms or the human on the phone hears a pause and starts talking over the agent. Barge-in detection, speculative TTS, and cutting a retrieval round-trip are where the engineering lives.
TakeSales / TakeFlow AI — multi-tenant voice & chat agent platform (Founder · technical lead)
A VAPI/Retell competitor built for the Brazilian market. Customers build, deploy and operate conversational agents wired into real phone lines, grounded in their own knowledge base.
| Latency | sub-900ms end-to-end voice loop |
| Retrieval | hybrid RAG (pgvector + keyword), tenant-isolated |
| Observability | Langfuse traces + cost, OpenTelemetry distributed tracing |
| Infra | AWS ECS Fargate · SQS · CloudFront · WAF · automated CI/CD |
| Design system | The Quiet Machine (in-house) |
Case study — a real-estate AI SDR agent: qualifies inbound leads and searches live property inventory via Elasticsearch during the conversation, with prompts tuned for Brazilian Portuguese.
Condo management SaaS — multi-tenant platform for building managers (Founder · technical lead)
Digitizes an entire regulated domain — resident communication, finance, bookings, incidents, documents and assemblies — with an AI layer that generates, summarizes and answers questions over meeting minutes via RAG.
- Multi-tenant architecture with per-condominium data separation and role-based authorization
- Translating quorums, cost apportionment and financial reporting into something non-technical users actually use
- TypeScript end to end · Fastify · Next.js · Supabase/pgvector · AWS
AI deals & coupons platform — crawler-to-search pipeline at scale (architected end to end)
crawlers → AWS SQS → workers → OpenAI enrichment → PostgreSQL → Fastify API → Next.js
- SEO as a first-class pillar: SPA → Next.js SSR/ISR migration, sitemaps, Schema.org structured data, dynamic metadata
- MongoDB → PostgreSQL (Prisma) migration for consistency and complex queries
- Price history, voting/comments, hybrid search, community spaces, affiliate monetization
Everything else in flight — 8 more projects
| Project | What it is | Status |
|---|---|---|
| Treinou.app | Fitness accountability app · ASO + store listing | Publishing |
| claude-kanban | Local kanban with markdown as source of truth; headless Claude Code, planner/scorer agents | Evolving → BYOM SaaS |
| Licitações AI | Public-tender opportunity discovery pipeline on real data | Running locally |
| Sereno | AI journaling / emotional wellbeing, reusing TakeFlow infra | Spec / MVP |
| Virtual Staging MVP | AI virtual staging for real estate | MVP |
| MCP Orchestration Gateway | Gateway architecture for MCP orchestration | Architecture |
| AI CMO | Internal multi-product social media management | Internal |
| Meeting-minutes AI | Condo minutes generation & Q&A (Skip Challenge hackathon) | Hackathon |
The honest breakdown — depth vs. familiarity
Daily driver, deep: TypeScript · Node.js + Fastify · React/Next.js (App Router, RSC, SSR/ISR/SSG) · PostgreSQL/Supabase + pgvector · AWS (ECS Fargate, Lambda, SQS, S3, CloudFront, WAF, RDS) · Docker · GitHub Actions
Applied AI, in production: LLM orchestration & tool calling · multi-agent workflows · hybrid RAG · Deepgram/ElevenLabs/Gemini Live · Twilio · Langfuse · OpenTelemetry · RAGAS · MCP · Claude Code headless
Strong: Python · React Native · Flutter · Kubernetes · Elasticsearch · MongoDB · Prisma · TailwindCSS/Shadcn · Zustand · React Query · WebSockets
Have shipped with, rustier: PHP · Java · Oracle · SQL Server · AngularJS · Appium/Selenium
Conversational AI lineage: Dialogflow & IBM Watson since 2019 → production LLMs since 2023. Not a 2024 pivot.
timeline
title 11+ years, from first line to production AI
2010 : First lines of code
2014 : Professional dev · CS @ UFU
2015 : Nexmob — web · mobile · Node
2018 : Atomic Agro — APIs, DevOps, data
2019 : Nexmob — low-code chatbot platform<br/>Dialogflow · Watson
2020 : Bittar Neurociência — Tech Lead<br/>LMS · CMS · CRM · live video at scale
2023 : Applied AI focus — LLMs in production
2024 : Founded Mobize<br/>Voice agent platform · condo SaaS
2026 : Voice agents · hybrid RAG · agentic CI
| 🥉 | 3rd place — Hackatri (Tribanco Hackathon) |
| 🏅 | 50 Best Apps of the Year (2015) |
| 🎖️ | Honorable mention ×2 — Brazilian Public School Math Olympiad |
| 🔭 | Honorable mention — Brazilian Astronomy & Astronautics Olympiad |
| ⚛️ | Senior React Engineer — Triplebyte Certified |
| 🎓 | BSc Computer Science, UFU · Postgrad in Generative AI Applications (2025–2026) |
♟️ Off the clock
Chess (competitive streak included), running and lifting, DIY and metalworking, film, games. The common thread is the same one in my engineering: I like systems with hard constraints and no room to hand-wave.
Structured specs first. I turn problems into dense documents before writing code — which is exactly why coding agents work well for me instead of flailing.
- Simple before clever. Clean, readable code that the next person can own.
- Performance work when it moves a real number, not for sport.
- Scalability designed in from day one; APIs that are boring to consume.
- Understand the business problem before the implementation.
- Direct, honest assessment — I'd rather be corrected than validated.
- Stack choices driven by delivery speed and cost, not hype.
