LLM applications and RAG
Assistants and search tools grounded in your own documents, with retrieval, embeddings and evaluation designed so answers stay accurate.
About
I design and ship AI systems built on large language models, and I care about the part that comes after the demo: making them reliable in production.
At Deloitte I architected the multi-agent orchestrator behind TaxBrain, a flagship platform that helps enterprise clients navigate Brazil's 2026 tax reform. At CEIA I lead an event-driven, multi-agent WhatsApp assistant running on AWS serverless.
My background is NLP research. I led the training of MED-LLM-BR, the first clinical LLM for Brazilian Portuguese, which received the 3rd Place Best Paper Award at BRACIS 2024.
- Based in
- Brazil, working remotely
- Focus
- LLMs, AI agents, RAG, MCP
- Languages
- Portuguese (native), English (highly proficient)
What I buildProblems I can take from idea to production.
Multi-agent systems
Agents that plan, call tools and hand work to each other, orchestrated with clear routing and control over prompts and behavior.
Fine-tuning and evaluation
Domain-adapted models trained on your data, and the benchmarks and metrics to prove they are better than the base model.
Production AI infrastructure
Scalable inference and event-driven pipelines on the cloud, from Kubernetes clusters to serverless queues and functions.
Skills
Languages
LLM and AI frameworks
Techniques
Cloud and infrastructure
Databases
Tools
Experience
-
May 2025 – Present
Senior GenAI Software Engineer
Deloitte, Disruption Office
Architected and developed the multi-agent orchestrator behind TaxBrain, one of Deloitte's flagship Tax assets: a high-investment platform that helps enterprise clients navigate Brazil's 2026 tax reform. Built the core chatbot with AI agents and RAG to retrieve and surface specific regulatory changes from the new legislation.
- Multi-agent orchestration: built the orchestration layer with LangGraph and Flask, running dynamic multi-agent workflows on Kubernetes (AWS EKS) for production-grade inference.
- Model Context Protocol: redesigned the orchestrator used across the Disruption Office around MCP, giving teams finer control over tool routing, prompt management and agent behavior.
- Invited talks: presented GenAI applications to PRODAM (São Paulo City IT Agency) and at the Jardim Digital event, representing Deloitte's AI innovation practice.
References
- Paulo Lima, LinkedIn
-
Feb 2025 – Present
AI Researcher
CEIA, Centro de Excelência em Inteligência Artificial
Leading the end-to-end development of a multi-agent WhatsApp chatbot, from the Meta WhatsApp Business API integration to production deployment.
- Event-driven architecture: designed the system on AWS Lambda, SQS FIFO queues and Google's Agent Development Kit, following message-bus patterns for reliable, ordered processing.
- Serverless infrastructure: owned the full setup on AWS, including webhooks, message routing and real-time agent orchestration.
References
- Anderson Carlos, LinkedIn
-
Feb 2024 – Feb 2025
NLP Data Scientist
Applied clinical NLP and LLMs to match patients with clinical trials, working on production pipelines that analyze eligibility criteria in medical notes.
- Clinical LLMs in production: integrated open-source clinical models (BioMistral, Meditron) into the pipeline, improving automated analysis of study criteria.
- Semantic search: fine-tuned embedding models, increasing retrieval accuracy for clinical trial matching.
- Evaluation and monitoring: built evaluation methods for model outputs and a real-time monitoring dashboard, making the pipeline more reliable.
- Com+Care: contributed to a multi-agent WhatsApp chatbot (LangGraph) that answers patients' questions about clinical studies and surfaces relevant ClinicalTrials.gov research.
-
Oct 2023 – Feb 2025
NLP Researcher
HAILab-PUCPR, Health Artificial Intelligence Lab
Led the development of LLMs specialized in Brazilian Portuguese clinical and biomedical text, producing the first LLM trained on Brazilian clinical data. Owned the data lifecycle, from cleaning and de-identifying patient records (LGPD-compliant) to training, evaluation and inference.
- Continued pre-training: domain-adapted LLaMA 2 7B on clinical and biomedical corpora with LLaMA-Factory and Hugging Face Accelerate, on Vertex AI from a single A100 to multi-GPU.
- Fine-tuning: applied LoRA, QLoRA and full fine-tuning (SFT) on biomedical question-answering datasets.
- Evaluation: benchmarked models on MMLU and custom clinical metrics, setting baselines for Portuguese clinical NLP. Published at BRACIS 2024, where the paper won 3rd Place Best Paper.
-
Sep 2022 – Feb 2024
Junior Data Scientist
Cadastra, SEO and Data team
Built data pipelines, scrapers and LLM-powered tools for the SEO practice of a digital marketing agency, used by enterprise clients such as Samsung, Vivara and Gerdau.
- LLM content generation: shipped a pipeline that feeds crawler and analytics data (Screaming Frog, Ahrefs, GA4) into the GPT API, delivered through an automated dashboard adopted by 32 clients. Extended it with LangChain agents for on-brand product descriptions and articles.
- RAG onboarding assistant: won the company hackathon with a Q&A assistant built on LangChain, ChromaDB, Streamlit and the ChatGPT API over internal onboarding documents.
- Cloud data collection: ran Selenium scrapers on AWS EC2 for Search Console, Google Trends, robots.txt and server logs, storing results in S3.
-
Feb 2019 – Apr 2022
Production Planner Analyst
Malta Brewery
Planned and monitored brewery production lines with the Head Brewer and Industrial Manager. This is where I first used data to drive operational decisions.
- Planning and analytics: built spreadsheet models and dashboards for weekly production planning and daily reporting, helping reduce production losses.
- Inventory and ERP: owned the monthly closing of raw materials and inputs, including reconciliation in the ERP system.
Publications
BRACIS 2024, Springer Nature
MED-LLM-BR: Developing Resource-Efficient Clinical LLMs for Brazilian Portuguese
Introduces the first clinical LLM for Brazilian Portuguese, trained on diverse clinical datasets with parameter-efficient fine-tuning (LoRA, QLoRA) and validated with a robust evaluation protocol. Received the 3rd Place Best Paper Award.
Achievements
-
2024
Best Paper Award, 3rd place, BRACIS 2024
Lead author of MED-LLM-BR, selected among hundreds of submissions at the Brazilian Conference on Intelligent Systems, Brazil's premier AI conference.
-
2024
Intel Software Innovator Program, Highlight of the Month
Recognized by Intel for building a quantized GGUF version of the clinical LLM with Alessandro de Oliveira, enabling efficient edge deployment of healthcare AI.
-
2025
Invited talks on Generative AI
Speaker for PRODAM, São Paulo's city IT agency, and at the Jardim Digital event, representing Deloitte's AI innovation practice.
-
2022–24
Company hackathon winner
First place with a RAG assistant that answers new hires' questions from internal onboarding documents.
Education
-
Postgraduate, Software Engineering: Automation and Innovation with Generative AI
Universidade Federal de Goiás (UFG), AKCIT/Embrapii. 2025 – present.
-
BTech, Big Data Technology and Analytics
Pontifícia Universidade Católica do Paraná (PUC-PR), Londrina. Sep 2020 – Apr 2023.
Courses and certificates
-
Deploying Machine Learning Models
Jun 2023. Serving models on AWS, GCP and Azure with MLflow, Kubeflow, Flask and Streamlit.
-
Fundamentals of Software Engineering
Oct 2022. Clean, tested and versioned Python; code review, logging and debugging.
-
Google Analytics Advanced
Oct 2022. Custom dimensions, segmentation, attribution and reporting.
-
Introduction to Statistics
Sep 2022. Probability, sampling, hypothesis testing and regression.
-
Python for Data Analysis 3.0
Jul 2022. pandas, NumPy and visualization for exploratory analysis.
-
Big Data Fundamentals 3.0
Jul 2022. Data lakes, data warehouses and the Hadoop/Spark ecosystem.
-
Introduction to Big Data
Dec 2021. Distributed processing with Apache Spark.
ProjectsEarlier machine learning work. More AI projects coming soon.
-
Cryptocurrency-Daily
End-to-end app for cryptocurrency price prediction, with a front end, a back end and a machine learning model for daily analysis and forecasts.
View on GitHub -
Customer Churn Prediction
Models that predict churn for a telecom company from customer demographics, account data and usage, so the company can act before customers leave.
View on GitHub -
Water Potability Prediction
Predicts whether water is safe to drink from physicochemical measurements such as pH, hardness and sulfate.
View on GitHub
“There is nothing outside of yourself that can ever enable you to get better, stronger, richer, quicker, or smarter. Everything is within. Seek nothing outside of yourself.”
Contact
Building something with LLMs or agents? Let's talk.
The fastest way to reach me is email or LinkedIn.