Applied AI Engineering

AI systems that move from prototype to production.

João Víctor López Matias builds deployment-minded AI systems across RAG, MCP, prediction pipelines, agent workflows, and product interfaces — connecting model capability to real business execution.

His work spans FI Group internal AI systems, earlier production computer-vision deployments, legal and retrieval research, and Speco: a control plane for specialist AI agents.

RAG systems
MCP servers
Agent workflows
Prediction pipelines
Computer vision
Speco founder

See systems Why this became Speco

Precedent Analysis Agent — grounded RAG microservice with vector store, source-cited answers over Lei do Bem precedent.

Systems, not demos

What João actually builds.

The common thread across João's work is not one model type. It is the ability to connect models, data, tools, evaluation, orchestration, and product surfaces into systems people can actually use.

Retrieval systems

Grounded workflows that retrieve domain-specific context, cite sources, and reduce hallucination in high-stakes business environments.

RAGvector DBsource grounding

Agentic tool execution

Agents that call tools, validate parameters, route requests, and return structured outputs instead of loose chatbot responses.

MCPtoolsschemas

Prediction pipelines

Feature extraction, embeddings, and model artifacts that turn raw reports into advisory intelligence and decision support.

embeddingsMLadvisory outputs

Productized AI

Interfaces, APIs, auth, deployment surfaces, and feedback loops that turn AI systems into real products.

UXAPIdeployment

Deployed Systems

Enterprise AI workflows, built end-to-end.

A selection of João's AI engineering work across grounded retrieval, predictive intelligence, MCP deployment, and agent orchestration.

System 01

FI Group · BR Lei do Bem

Precedent Analysis Agent.

Grounded legal-intelligence workflow for cited answers over domain-specific precedent.

Problem

High-context legal and policy workflows require answers that are not just fluent, but grounded in evidence.

System

An agent workflow connected to a RAG microservice, vector database, and LLM layer for retrieving Lei do Bem precedent material and returning answers with sources.

Why it matters

It shows João's ability to design AI systems where source grounding, retrieval quality, and production usability matter more than generic chatbot behavior.

RAG microservicevector DBsource-backed outputdomain retrievalgrounded answers

System 02

FI Group · BR Lei do Bem

Technical Intelligence Prediction Agent.

Prediction-oriented AI workflow for structured advisory outputs.

Problem

Technical reports contain signals that need to be extracted, embedded, and transformed into decision-support outputs.

System

A pipeline combining report extraction, section embeddings, ML artifacts, prediction outputs, recommendation probability, top motives, and advisory metadata.

Why it matters

It shows João's ability to move beyond text generation into AI systems that combine extraction, embeddings, prediction, and structured outputs.

prediction pipelineembeddingsML artifactsadvisory outputstructured features

System 03

FI Group · MCP-Brasil + MarIA

MCP-Brasil + MarIA Deployment.

Tool-grounded agent deployment connecting APIs, internal context, and structured responses.

Problem

Enterprise agents need controlled access to tools, official data, internal context, and reliable response structures.

System

An end-to-end architecture where MarIA routes work through MCP-Brasil, accesses public Brazil APIs and FI Group RAG enrichment, then returns validated tool-grounded responses.

Why it matters

It shows João's deployment-minded AI engineering: MCP servers, tool execution, orchestration, structured outputs, and business-ready agent workflows.

MCP servertool executionAPI integrationstructured responsesdeployment architecture

Architecture Direction

The next layer: orchestrated specialist agents.

Alongside deployed systems, João has also designed architecture directions for multi-agent coordination inside MarIA: routing requests across specialized agents, combining outputs, and returning one unified response.

Forward-looking architecture

FI Group Brazil · MarIA

Multi-Agent A2A Orchestration.

A shared orchestration layer routes requests to the right specialist agent — precedent analysis, prediction, MCP-Brasil, or future modules — then merges the outputs into a unified response.

multi-agent routingA2A callsresponse mergeshared contextspecialist agents

Why this became Speco

After building internal AI systems, the pattern became obvious.

Across retrieval, prediction, MCP, and agent workflows, the same problem kept appearing: teams do not just need access to foundation models. They need an easier way to turn their own knowledge into reliable, deployed AI agents.

Speco is João's product answer to that pattern: Vercel for training and deploying AI agents. A company uploads its data, creates a specialist agent, tests it in the Playground, and deploys it through API and MCP — without managing prompts, evals, infrastructure, model optimization, or MCP server setup. Train and deploy AI agents from your data.

Visit Speco Back to homepage

Step 01

Knowledge becomes datasets

Raw documents become structured examples, traceable sources, and inspectable training material.

Step 02

Quality comes before deployment

Eval-first workflows make specialist behavior measurable before teams rely on it.

Step 03

Specialists become infrastructure

Deployed agents become API and MCP surfaces for downstream products and workflows.

Earlier Production Work

Computer vision systems in production contexts.

Before João's current focus on agent systems and AI infrastructure, he also shipped applied computer-vision systems across use cases such as PPE detection, cellphone usage, posture, hand-washing, and sentiment-related classification.

Earlier foundation · 2024–2025

computer visionPPE detectionposturehand-washingproduction AI

Engineering Edge

A rare combination of AI depth, product taste, and deployment speed.

What distinguishes João's applied AI work — six things, named explicitly.

Systems thinking

João connects models, data, retrieval, tools, UX, and deployment into one working system.

Grounded AI instincts

He designs for source grounding, structured outputs, and traceable behavior instead of vague chatbot magic.

Agent-native architecture

His work spans MCP, tool execution, orchestration, and specialist-agent workflows.

Production bias

The focus is not demos. It is systems that can be inspected, shipped, monitored, and used.

Founder-level product sense

Speco turns the patterns from applied AI work into a product customers can understand and use.

Fast AI-assisted execution

João builds with modern AI engineering workflows, using tools like Claude Code and Cursor to move from idea to polished product fast.

From models to systems

This is the engineering behind Speco.

João's applied AI work is less about isolated models and more about systems that can be deployed, inspected, orchestrated, and used by real teams — from retrieval and prediction to MCP and multi-agent workflows.

Back to homepage Visit Speco