Selected work

LLM fine-tuning infrastructure for Mexican Spanish political analysis.

Specialized LLMs, produced by a reproducible fine-tuning and evaluation pipeline. Built for empirical comparison of base models and prompts.

A legal research workbench for the persistent challenge of jurisprudence search, with factuality at the centre of AI-assisted work.

Jurisprudence search and advanced RAG bring sources into the drafting process, with tools designed to reduce hallucinations and keep claims open to verification.

A modular fact-checking toolkit designed to ease the workload of journalists facing a growing tide of misinformation.

Automated claim verification brings claim detection, evidence retrieval and assessment into a transparent workflow. Its modular stages are designed to reduce repetitive work and keep sources and reasoning open to inspection, giving journalists more room for the investigations that need human judgment.

AI-powered transcription and analysis pipeline for political focus group research.

A full audio-to-insight pipeline replacing error-prone transcription and hours of note-taking per project. Speaker diarization, transcription, theme classification, and integrated reporting — all surfaced through a tooling layer that researchers actually use.

Automated participant screening system for qualitative recruitment.

Automated intake pipeline that replaces hours of manual data entry per project. Combines OCR, fraud detection, and socioeconomic classification — so analysts focus on the decisions, not the paperwork.

Turns rough notes into an editable, on-brand PowerPoint at the press of a button.

The brand template is codified once, so every slide comes out compliant by construction. Content lives in plain text and the slides regenerate from it, versionable and reproducible like code. The output is a standard .pptx that anyone can edit, with no platform in the loop.

Python scraper and parser for Brazilian Supreme Court (STF) case data.

Typer-based CLI with three cache-first stages — scrape, download, extract — supporting heavily sharded runs with proxy rotation, feeding a DuckDB warehouse. Multiple OCR backends, including self-hosted Tesseract on fly.io for cheap inference.

CLI for an AI agent to keep a Zotero library in order.

Checks tags, metadata, and collections; enriches metadata from Crossref and arXiv; compresses PDFs. Implements a validation loop with backups and write-ahead logs — making it safe to let an AI agent touch the database.