DATA · ML · AI PLATFORM
I build data platforms and AI systems that teams can inspect, reproduce and operate: lakehouses, ML pipelines, retrieval systems and the models behind them.
TRY IT
Each demo covers one area teams hire for, from generative AI to data governance, running on the public Brazilian datasets published on Kaggle. They describe what the records say; they are not legal advice or a supplier recommendation.
Generative AI · RAG
A self-hosted language model with no refusal layer, grounded in knowledge bases: it searches procurement notices, contracts and education spending, writes read-only SQL when the answer is a number, cites every record and keeps the thread of a conversation.
Open the chat
Vector search · Semantic retrieval
1.25 million procurement notices embedded with a self-hosted model and indexed in pgvector: compare keyword search with search by meaning, side by side, next to the aggregate dashboard.
Search notices
BI · Analytics
An analytics dashboard over municipal education spending: investment per student by year, region and state, built on modelled semantic tables.
Open the dashboard
Data lake · Catalogue · Governance
The catalogue of a public data lake: 428 tables and 33,891 columns from raw to analytics layers, with source, lineage, join keys and the privacy rules applied, searchable in the browser.
Browse the map
OPEN SOURCE
Every demo and research project above, grouped by track, with its repository, published data and stack.
Eleven Brazilian public sources, from the school census to procurement, released as raw, trusted and analytics layers with contracts, a privacy gate and per-file hashes.
31 Kaggle datasets, 428 tables and about 4 billion rows of Parquet. Every file is listed in a SHA-256 manifest, and a clean download is verified against it.
A specification-driven cost assessment that turns a narrow metadata access boundary into auditable findings, a report and a deck.
Municipality-level anomaly triage on public education spending, with lineage and a drift gate that refuses to score a shifted batch.
v0.2.0 scored the 2022 batch with 96 review signals; the drift gate blocked the 2023 batch at a spread ratio of 1.384.
Outputs are review signals. There are no labels, so no accuracy is claimed.
Transparent retrieval and ranking of historical procurement notices, with every score explained.
v0.2.0 ranks notices from the pinned PNCP release and verifies its hash before use.
Offline evaluation uses synthetic profiles, so it measures constraint adherence, not user relevance.
A cited research chat over three public bases: text and vector retrieval for notices, read-only SQL for contracts and spending, and gates that run before any model call.
1,250,335 notice embeddings in pgvector. The 2026-10-01 browser session answered 10 of 10 scripted questions.
Numeric questions across all bases can still fall back to text retrieval; a question router is in progress.
The 27B Qwen model behind the chat, served from one rented GPU through an OpenAI-compatible API.
NVFP4 weights on vLLM with speculative decoding: from 4.4 to about 34 tokens per second on the same GPU.
The abliterated checkpoint is a third-party release; I did not retrain it.
A retrieval reference with evaluation, source citations and redacted traces.
Redis-coordinated workers with idempotent requests, Kubernetes manifests, Terraform and recovery tests.
48 requests at concurrency 6: 61.89 req/s, p95 147.95 ms. Worker recovery proven on a local kind cluster.
Deterministic model stub on one host; not a capacity claim. Cloud not validated.
CAREER
Ten employers across retail, banking, payments, credit and education. Select one to see what I did there.
DEC 2024 – NOW · 1 YR 10 MO
Data and AI engineering · Retail
Data and AI engineering for one of Brazil's largest pharmacy chains: pipelines, models and AI agents delivered in one engineering flow, with governance and data quality built in.
First year a tool appears in a role above.
2014
Java · C++ · Python · MQTT
2016
Delphi · Oracle · PL/SQL
2017
Spark · Airflow · Azure DevOps · SAP HANA
2019
Node.js · React · Docker · AWS · SSIS · SQL Server
2021
GCP · BigQuery · Hadoop · Hive
2022
Vertex AI · Dataproc
2023
Databricks · PySpark · MLflow · Delta Lake
2024
dbt · GitLab CI · LLM · RAG
Client work is described by sector and stack only. Names and results stay with the clients.
GROCERY E-COMMERCE
Product recommendations for a European online grocer, chosen in two stages: first the category, then the brand and pack size.
AWS Glue · SageMaker Pipelines · DeepFM · DIN · Terraform
ENTERTAINMENT
Concession recommendations for registered and anonymous customers, with an A/B design.
Azure ML · Microsoft Fabric · Azure DevOps
FINANCE
A multi-agent assistant whose every number comes from a bronze, silver and gold lake, never from the prompt.
AWS Glue · Athena · SageMaker · LLM agents
MINING AND BULK MATERIALS
Stockpile volume from drone imagery: photogrammetry, ground fitting and point-cloud segmentation in a container job.
OpenDroneMap · ECS Fargate · Open3D · DBSCAN
CREDIT
A self-hosted data lake and risk engine for receivables financing, scoring financial risk and bad faith on separate axes.
ClickHouse · MinIO · Airflow · Metabase · FastAPI
EDUCATION
Public-data and product lakes for education companies: raw, trusted and semantic zones, daily orchestration and BI.
BigQuery · Airflow · Metabase · Power BI
LEGACY MODERNISATION
A specification-driven transpiler that moves COBOL, Delphi and SAS code to Python and Databricks.
LLM · SDD · Databricks
Engineering first, then software, then data and AI: each degree matches a step in the career above.
Bachelor's degree
Instituto Federal de Goiás (IFG)
Electronics, control systems, embedded programming and industrial automation: the base of the early work in embedded software and automation.
Technologist degree
Universidade Norte do Paraná (UNOPAR)
Software engineering, databases, systems analysis and application development.
Specialisation
Universidade Norte do Paraná (UNOPAR)
Data pipelines, data modelling, warehouses and big data processing.
Specialisation
Centro Universitário UNYLEYA
Machine learning models, their evaluation and applied artificial intelligence.
WRITING
CASE NOTES · 8 MIN
What memory-bandwidth arithmetic, speculative decoding, a paired benchmark and an audited grader taught me about serving an open model at interactive speed
CASE NOTES · 4 MIN
One rule, gates that run before any model call, and what broke on the first live runs
THE PUBLIC DATA STACK · 6 MIN
A tutorial: load a public-data lake into pgvector, index it three ways, give each reader its own limits, and keep it fast on a small server