Projects organized by solution type, stack and technical focus

Generative-AI (RAG) assistant for public-sector support

I designed and shipped a production generative-AI (RAG) assistant supporting 20+ scholarship, benefit and accountability programs in the federal public sector, across two channels: an assisted-consultation app and an Outlook reply-drafting assistant.

What I built: a full RAG architecture with a curated-answer layer (anti-hallucination), a curation queue, automatic routing, per-program observability and hybrid search with model fallback.

Invoice audit with PySpark, Databricks and Genie

Big data analytical engineering to investigate transaction inconsistencies, with analytical tables, PySpark queries and a dashboard.

Impact: around 20% of inconsistencies detected and audit time cut in half.

Contract reading and payment calendar with AI

Extraction of financial clauses from PDF contracts, expected payment calendar construction and divergence monitoring.

Impact: contracts paid on the correct dates, generating direct financial savings and tighter internal audit control.

RAG NLP SQL with LangChain, OpenAI and SQLite

Natural language questions over a relational database, combining semantic schema retrieval with SQL generation and execution.

MCP Docs Assistant with FastMCP and BM25 search

Read-only MCP server for querying local documentation, exposing resources, tools and prompts to MCP-compatible clients.

Invoice audit with PySpark, Databricks and Genie

Analytical tables and PySpark queries built in Databricks notebooks to investigate transaction issues affecting invoice analysis.

• Goal: structure a reliable view of inconsistencies in high-volume data.
• Delivery: analytical tables, PySpark queries and views by inconsistency type.
• Stack: PySpark, Databricks notebooks, analytical modeling, Genie and dashboard layer.
• Technical highlight: big data analytics engineering, data quality and automated refresh.

Bank transaction audit with Random Forest

Machine learning project focused on prioritizing inconsistencies and risk signals in transactional data to support audit and analysis.

• Goal: classify risk signals and prioritize transactions for review.
• Delivery: supervised classification with tabular preparation and derived features.
• Stack: Python, pandas, scikit-learn, matplotlib, pyarrow, joblib and unittest.
• Technical highlight: supervised tabular ML and explainability for analytical support.

Outlier Detection Lab for inconsistencies and anomalies

Outlier and anomaly detection lab on large datasets to identify extremes, unlikely combinations and cases worth manual review.

• Goal: identify anomalies and extreme cases for review.
• Delivery: comparison of statistical and unsupervised approaches.
• Stack: Python, pandas, scikit-learn, robust statistics and unsupervised ML.
• Technical highlight: combining statistical methods and unsupervised models for audit support.

CadÚnico profile analytics

Project inspired by sampled public microdata to analyze income, registration status, family vulnerability and territorial prioritization.

• Goal: transform social microdata into territorial and managerial analysis.
• Delivery: vulnerability indicators, profile views and territorial analytics.
• Stack: Python, pandas, numpy, Streamlit and Plotly.
• Technical highlight: social indicators and territorial prioritization.

Contract reading, payment calendar and escalation with AI

Solution for reading contract PDFs, extracting financial clauses, building expected payment calendars and tracking divergences.

• Goal: transform unstructured documents into a monitorable financial flow.
• Delivery: contract reading, clause extraction and expected payment calendar generation.
• Stack: Python, regex, PDF processing, analytical architecture and agent layer.
• Technical highlight: contract extraction, financial reconciliation and workflow automation.

Technical Request Document Assistant

Integrated flow for technical request reading, structured PDF field extraction and related reference retrieval.

• Goal: structure document reading and reference lookup in a single interface.
• Delivery: structured PDF extraction and reference retrieval in one app.
• Stack: Python, reportlab, pypdf, pandas, scikit-learn and Streamlit.
• Technical highlight: structured extraction, semantic retrieval and document assistant design.

Engineering Document Consistency AI

Pipeline for clause extraction, semantic search, inconsistency detection and human review in a dashboard.

• Goal: compare documents, retrieve relevant passages and support inconsistency review.
• Delivery: document comparison and semantic retrieval of related passages.
• Stack: Python, reportlab, pypdf, pandas, scikit-learn, Streamlit and Plotly.
• Technical highlight: document governance, semantic comparison and review workflow.

Political and Economic News Intelligence Dashboard

Web scraping project using `newspaper3k` to collect news, structure an analytical dataset, apply NLP and publish an interactive dashboard.

• Goal: collect, enrich and visualize news content in an analytical structure.
• Delivery: scraping, NLP and dashboard with thematic and entity views.
• Stack: Python, newspaper3k, pandas, spaCy, Streamlit and Plotly.
• Technical highlight: automated collection, NLP over news and executive visualization.

Generative-AI (RAG) assistant for public-sector support

I designed and shipped, end to end, a production Retrieval-Augmented Generation system supporting 20+ scholarship, benefit and accountability programs in the federal public sector. It works across two channels: an assisted-consultation app for the team and an Outlook-integrated assistant that summarizes messages and drafts replies, always reviewed by a person before sending.

• What I built: the full RAG architecture, from data to interface, focused on accuracy and reliability.
• Anti-hallucination: a curated-answer layer before generation and a curation queue (review and publish), ensuring accuracy on amounts, deadlines and legislation.
• Routing and human-in-the-loop: automatic routing to the responsible area and mandatory review of drafts before sending.
• Observability: per-program performance dashboards, interaction telemetry and unanswered-question mapping for continuous improvement.
• Retrieval: hybrid vector search (dense and sparse), resilient model fallback with circuit breaker and confidence scoring.
• Platform: role-based access control, an internal knowledge base and a team chat.
• Stack: Python, Streamlit, Databricks (Apps, Vector Search, Delta Lake, SQL Warehouse), BGE-large embeddings and LLMs with fallback.

RAG NLP SQL with LangChain, OpenAI and SQLite

Python application that answers natural language questions over a SQL database by combining schema-aware retrieval and SQL generation.

• Goal: allow natural language analytics over relational data.
• Delivery: app for question answering over SQL with semantic context and assisted query generation.
• Stack: Python, LangChain, SQLAlchemy, SQLite, Streamlit and BM25Retriever.
• Technical highlight: RAG applied beyond documents, improving structured navigation.

Search Performance Assistant for retrieval evaluation

Python application to study document retrieval with TF-IDF, vector indexing, fallback behavior and evidence-based assistant responses.

• Goal: explore retrieval in a transparent and comparable way.
• Delivery: ingestion, indexing, retrieval and visual explanation of search results.
• Stack: Python, scikit-learn, TF-IDF, FAISS, cosine similarity, Tkinter and unittest.
• Technical highlight: retrieval evaluation and traceable ranking explanation.

Release Notes Generation Assistant

Python app for assisted release note generation based on release context, pull requests, similarity retrieval and evaluation metrics.

• Goal: organize release context and PRs for assisted note generation.
• Delivery: release note generation with retrieval and thematic rules.
• Stack: Python, scikit-learn, TF-IDF, cosine similarity, Tkinter and unittest.
• Technical highlight: product-oriented retrieval pipeline with reproducible evaluation.

MCP Docs Assistant

Read-only MCP server for local markdown documentation with resources, tools and prompt support.

• Goal: expose local documentation in a format consumable by MCP clients.
• Delivery: read-only MCP server with catalog, search and retrieval.
• Stack: Python, FastMCP, rank-bm25, frontmatter parsing and markdown.
• Technical highlight: MCP design, BM25 retrieval and agent-ready documentation access.

MCP SQL Analytics Server

MCP server for SQL analytics exposing structured tools for schema inspection and analytical querying.

• Goal: enable structured data exploration by agents in a controlled environment.
• Delivery: MCP tools for inspection and SQL analytics.
• Stack: Python, MCP, SQL analytics and modular tool design.
• Technical highlight: agent-oriented data access and tool-based analytics.

Curriculo Site built with Codex and vibe coding

Personal website and technical portfolio with bilingual pages, booking pages and production publishing on a custom domain.

• Goal: consolidate a professional presence in a custom domain.
• Delivery: bilingual site with home, portfolio, booking flow and production deployment.
• Stack: HTML, CSS, JavaScript, GitHub, Vercel and manual technical content curation.
• Technical highlight: AI-assisted prototyping, content structuring and shipping.