Nicolas Santos

Data Engineer · Applied AI Developer

Computer Science @ UNIFOR  •  Fortaleza, Brazil

I build data and AI systems that run end to end — from raw public datasets to the interface someone actually uses. Most of my work sits at the intersection of data engineering and local-first LLM applications: RAG pipelines over clinical documents, natural-language interfaces to epidemiological databases, and desktop AI tools that keep inference on the user's machine instead of in someone else's cloud.

The recurring theme across my projects is privacy-preserving AI on real public-health data — systems that stay useful without shipping sensitive records anywhere.

Databricks Student Fellows

The Databricks Student Fellows program recognizes students building with the modern data and AI stack. Out of the whole cohort, five fellows were highlighted by the program — and I'm one of them.

_{The five featured Databricks Student Fellows.}

A conversational data analyst for Brazil's Mortality Information System (SIM/DATASUS). You ask a demographic or epidemiological question in plain language; the system generates Python, executes it inside an isolated Docker sandbox, and returns charts with an executive summary.

The hard parts weren't the LLM — they were the data engineering around it: autonomously downloading, caching and filtering DATASUS parquet files, translating cryptic CID-10 codes into readable diagnoses, and choosing a chart type that doesn't collapse under a long tail of categories.

Python Qwen2.5-Coder PySUS Docker Plotly

Pipeline Unique for Reading and Epidemiological Segmentation.

Clinical guidelines run to hundreds of pages, and healthcare data can't be shipped to a third-party API. P.U.L.S.E. is a fully local RAG system that ingests clinical PDFs and CSVs and answers questions against them — no data leaves the machine, LGPD-compliant by construction, zero inference cost.

Built on a Medallion-style architecture adapted for AI workflows: Bronze ingestion → Silver chunking and cleaning → Gold embedding and vector persistence → RAG retrieval.

Python LangChain ChromaDB Ollama (Phi-3) Pandas Streamlit

A desktop note-taking app where both the notes and the model live on your machine. A llama.cpp sidecar is spawned lazily, streams over loopback HTTP, and shuts itself down when idle. Inline AI actions on selected text, a chat panel grounded in the current note, and full-text search via SQLite FTS5 with BM25 ranking.

It also profiles the host's available RAM on first run and recommends a model that will actually fit.

Rust Tauri 2 React TypeScript SQLite (FTS5) llama.cpp Qwen3

Languages   Python · TypeScript · JavaScript · SQL · Rust · Java · R

Data & AI   Pandas · NumPy · LangChain · ChromaDB · Ollama · llama.cpp · Jupyter · Plotly · Streamlit

Platform   Docker · SQLite · Git · Tauri · React