Milene Martins · Artificial Intelligence Engineer
AI happens when engineering holds up the model.
I build production Artificial Intelligence systems, focused on GenAI, multi-agent architectures and LLM-based applications.
01Perspective
Building AI is not just making the model answer. It is making the system work in the real world.
I work at the boundary between software engineering, data and product: deciding where the model belongs, where it does not, how to measure whether it works, and what happens when it fails.
02How I think AI
Engineering decisions that hold a system up in production.
Production over prototype.
A demo proves it is possible. Production proves it is reliable. Failures, cost, latency and operations are part of the design from the start.
Deterministic where necessary.
LLMs where there is language and ambiguity. Explicit rules where there is policy, risk or obligation. Not every decision needs a model making it.
Observable by design.
If you cannot see what the system did and why, you cannot improve it. Tracing and evaluation are born with the first version.
Business impact matters.
An AI system is measured by what changes in the operation, not by the sophistication of its architecture.
03Experience
Where and how I applied this.
Corporate AI platform: conversational experiences built on LLMs and multi-agent architectures for digital channels. Technical Owner of strategic initiatives in the collections domain, from architecture to production.
- Technical Owner of strategic initiatives: from defining the architecture and business rules to deployment and continuous evolution in production.
- Conversational capabilities within a multi-agent architecture: agent behavior, conversational flows and business rules.
- Software engineering applied to LLMs, integrating APIs, distributed services and components of the corporate AI platform.
- Observability, traceability and continuous evaluation with Langfuse, Datadog and LLM-as-a-Judge pipelines.
- Platform reliability: hallucination mitigation, context management, response validation and handling of operational failures.
- Safe rollout with feature flags, staging environments and test automation, in collaboration with Product, Engineering and Business.
- 28.2%
- negotiation promises
- ~R$ 540k
- recovered in two months
- Award
- part of the winning project of an internal AI award
Deploying and operating LLM applications, and the data engineering that feeds them, across Azure, GCP and Databricks.
- Deployment and supervision of GPT models in production, with adjustments to keep them compliant with requirements and up-to-date data.
- Prompt engineering to improve the performance of generative models.
- Research into new LLM applications, such as chatbots and intelligent search, with Microsoft Azure and GCP.
- Extraction, processing, insertion and vectorization of metadata for AI model training.
- SQL and PySpark pipelines and queries, moving data between BigQuery (GCP) and Databricks in a layered architecture (raw, trusted and refined).
- Technical documentation, diagrams and presentations; agile sprints; Git, GitHub and Azure DevOps with Pull Request reviews.
Backend development in Python and Django, focused on code quality, testing and continuous delivery.
- Backend applications in Python and Django, following clean code, PEP 8 and linting, with a MySQL relational database.
- TDD to prevent regressions and ensure quality and security, plus integration tests of the routes (CRUD) via Postman.
- CI/CD on GitLab and Git versioning, using Conventional Commits.
- Log lookup on GCP and monitoring with Grafana.
- Merge request reviews with front-end context (JavaScript, React, HTML and CSS).
04Focus areas
The territory I work in.
Generative AI
From use case to production system, with the model as one part and not the product.
Multi-Agent Systems
Agents with clear roles, limits and responsibilities, coordinated within business rules.
LLMs
Model, prompt and context choices driven by the problem, cost and latency.
RAG & Vector Search
Retrieval that brings the right context, measured rather than assumed.
MCP & Tool Calling
Models that act on real systems through well-defined interfaces.
AI Observability & Evaluation
Traces, metrics and continuous evaluation to know when the system fails, and why.
Python / FastAPI
The service foundation: clear, testable APIs ready to operate.
AI System Design
Boundaries, flows and trade-offs designed before the first prompt.
05Education
Academic background and certifications.
Education
- Postgraduate Degree in Machine Learning EngineeringFIAP · 2025 – 2026
- Technologist Degree in Systems Analysis and DevelopmentUNIBTA · 2020 – 2023
Certifications
- AI Specialist Career TrackAlura · Jul 2026
- Generative AI FundamentalsDatabricks · Nov 2024
- Databricks for Data EngineeringDatabricks · Nov 2024
- Microsoft Certified: Azure AI FundamentalsMicrosoft · Jun 2024
06Contact
Milene Martins
Artificial Intelligence Engineer