Building with Claude
Too many things !
Six side projects, one AI assistant
There’s a particular kind of weekend energy that comes from having too many half-formed ideas and just enough time to turn them into something real. Over the past several months I’ve been working with NanoClaw, the an AI agent system I use for research, writing, and code on a handful of projects that range from geopolitics to grammar drills. None of these were commissioned.
learn-bangla
Bengali is spoken by ~230 million people and almost never taught in European schools. I wanted to learn it. Rather than relying on Duolingo’s gamified drip, I built a small dataset of sentence pairs and a simple flashcard site to practice against. NanoClaw helped structure the sentence corpus and sanity-check the transliterations. The result is minimal — a static site, a CSV, and a reminder that language learning is mostly about showing up every day with something to look at.
FIFA2026
The 2026 World Cup is the first with 48 teams, which means more matches, more upsets, and more interesting odds to model. I built a Poisson-based prediction tracker: de-vig the bookmaker odds into implied win/draw/loss probabilities, convert those to attack/defence strength parameters, simulate scoreline matrices, and track how predictions age against actual results. NanoClaw handles the nightly result updates and helps write the Poisson predictions for upcoming fixtures. The standings, predictions, and scoring files update automatically. It’s not going to beat the market — but it’s a good way to stay honest about uncertainty.
idf-datacenters
A more infrastructure-flavoured project: where in Île-de-France could you actually site a large datacenter (≥20 MW) given the electrical grid as it stands? I pulled RTE and Enedis substation data, overlaid OpenStreetMap land-use layers, ran a Linear Optimal Power Flow model to estimate headroom, and scored every plausible substation on seven criteria — grid capacity, proximity to fibre, flood risk, land availability, and a few others — out of 103 points. NanoClaw helped with the LOPF setup and with writing the analysis narrative. The output is a ranked list of substations with reasoning. Useful if you’re doing infrastructure foresight in the Paris basin; niche otherwise.
OrmuzOsint
The Strait of Hormuz is one of the most watched — and most actively obscured — maritime chokepoints in the world. This toolkit fuses Sentinel-1 SAR imagery with AIS vessel tracking to surface “dark” ships: vessels that appear in the radar data but have switched off their transponders. Everything runs on free data sources. NanoClaw helped with the fusion logic and with writing the documentation. The methodological point is simple but underappreciated: SAR doesn’t care if you turn off your AIS. The gap between what the radar sees and what the AIS reports is the interesting signal.
Resume_RDF_update
I got tired of having a CV that was a flat document with no queryable structure. This project parses CVs into Turtle RDF knowledge graphs using the ResumeRDF ontology — skills, roles, projects, organisations, time periods, all as linked data. There’s a CLI for batch processing and a Streamlit app for exploration. Once your portfolio is a graph, you can ask questions like “which skills appear across the most projects” or “what’s the overlap between two people’s experience” — questions a PDF simply can’t answer. NanoClaw helped build out the Streamlit interface and test the ontology mappings.
ITS-C5
The most embedded-systems project of the set. ITS-G5 is the 5.9 GHz radio standard used for vehicle-to-vehicle and vehicle-to-infrastructure communication in Europe. Most receivers are expensive or closed. This is firmware for the ESP32-C5 on a Seeed XIAO board: it captures ITS-G5 broadcast frames and streams them as newline-delimited JSON over USB serial. No network stack, no cloud dependency — just radio frames to your laptop. NanoClaw helped strip the firmware down and write the frame parsing logic. The goal is a cheap, open receiver for anyone doing ITS research or monitoring.
On working this way
What I notice across these projects is that they share a shape: a question that could be answered computationally, a data source that’s publicly available but annoying to wrangle, and a gap between “interesting idea” and “working prototype” that used to be a weekend of glue code.
The thinking is still mine (part of, the shaping at least =) . The itch is still mine. The code just arrives faster.
None of these projects are finished. All of them taught me something. That’s about the right ratio.
