FPFarina Pontejos
GitHubGet in touch
Selected work

Five projects

Each entry says what it was, what I did and what came of it. Where a figure is missing, it's covered by a confidentiality agreement or nobody measured it; I leave the field out instead of estimating.

Full-stack2026–now · Freelance data scientist · DareData

Cable engineering platform

Design, geometry and bill-of-materials tooling for an industrial client

Client work delivered through DareData: a family of services that turn cable specifications into geometry, simulations and bills of materials, with a React front end on top. I work across the data model, the API and the test suite. The client and the numbers are covered by a confidentiality agreement; ask me and I’ll talk through the architecture in conversation.

PythonFastAPITypeScriptReactPostgresDocker
Repositories
8
Client
Confidential
Data / ML2026 · Sole author

claude-logs

An event spine for AI coding-session logs

An ETL that reads Claude Code’s JSONL transcripts into a DuckDB event spine (turns, tool calls, agent spawns, messages), plus a read-only FastAPI and React inspector for browsing sessions. Built to answer questions about how I actually work with agents: which tools fail, how long turns take, what I type versus what gets injected.

PythonDuckDBFastAPIReactDocker
Sessions ingested
1,445
Events
1.3M
Hand-typed prompts
9,092
Open source2022–2025 · Adjunct lecturer · NOVA IMS

Data Mining course materials

Four editions of practical classes, all public

Notebooks and exercises for the practical sessions of the MSc Data Mining course at NOVA IMS, from 2022/23 to 2025/26, plus a Python crash course and a mock repository for teaching git. Everything is on GitHub under fpontejos, reused by students after the semester ends.

Pythonscikit-learnJupyterGit
Editions
4
GitHub stars
45
Data / ML2021–2026 · MSc thesis · PhD proposal

Bug-report triage research

From self-organising maps to LLM agents that judge taxonomies

My MSc thesis at NOVA IMS used self-organising maps to triage software bug reports; the code is public as sombr-thesis. The PhD proposal that followed asked whether LLM agents could evaluate automatically generated taxonomies: how you judge a classification nobody wrote by hand. It was an interesting question, but industry turned out to be even more so, and I set it aside in 2026 to work full-time.

PythonSelf-organising mapsText miningLLM agentsQuarto
MSc thesis
2023
PhD proposal
2023
Full-stack2026 · Sole author

pointy.dev

This site, with the numbers mined from my own git history

Static Astro site with every piece of copy in typed content collections and no client-side JavaScript. The design started as a Claude Design export and was ported to native components. The skills and stats on these pages come from a pipeline that reads my repositories and coding-session logs: aggregates only; nothing covered by a confidentiality agreement leaves the machine.

AstroTypeScriptContent collectionsPythonDuckDB
Client-side JS
0 kB
Dependencies
1

Client work is described without names or numbers because of confidentiality agreements. Ask me and I'll walk through the architecture in conversation.

What are we building next?

Tell me about the data and what it needs to do, and I'll let you know whether I'm the right person for it.

Message on LinkedInBrowse the code