An OpenWebUI ↔ OpenCode bridge for querying a personal markdown wiki with an LLM. Deliberately no RAG and no embeddings: the model reads the actual files, orienting itself from a root index.md (based on LLM Wiki idea).

An AI system that matches incoming bank transfers against customer insurance policies, handling the messy free-text references humans put in payment descriptions. Reduced manual reconciliation time by roughly 80%.

An MCP server that exposes a SiYuan knowledge base to LLM clients such as Claude Code, down to individual notebooks, documents and blocks, so an agent can read and write personal notes as a first-class tool rather than through copy-paste.

A scalable scraping architecture built on AI agents and a custom MCP server. Agents handle the parts that change: page structure, selectors, edge cases. The surrounding workflow stays deterministic, so runs are reproducible and failures are debuggable instead of mysterious.

Improved multimodal LLM performance using a C-abstractor adapter between the vision encoder and the language model, trading a small amount of spatial detail for a large reduction in visual token count.

Android app that mirrors an entire Navidrome library (or any Subsonic-compatible server) into a local folder any player can read. Playlists export to .m3u8 with relative paths, and starred tracks sync both ways.

Not rsync: Subsonic only speaks HTTP, so the sync is idempotent logic on top of the API, index the server, diff against the folder, and pull only what is missing or changed via the download endpoint, which returns the original file rather than a transcode.

Downloads YouTube audio as MP3, writes the metadata, and files it into a personal library. Driven from a Telegram bot, so adding a track is one message from a phone: send a link and it is either delivered back as a file or saved straight to the server, depending on who is asking.

Two machines, in two different houses, running about twenty services behind a reverse proxy and joined by a zero trust network rather than an open port. Photos, music, audiobooks, ebooks, notes and files all live there rather than in someone else’s cloud, and a dashboard tells me at a glance what is healthy and what is not.

Part of it is stuff I wrote: a personal wallet app with its own API and Postgres database, a Telegram bot that reports how much the solar panels are producing, and a few smaller things that exist only because it was quicker to build them than to look for them.

It is also where anything new gets tried first. If a tool survives a few weeks of me actually depending on it, it stays.

• Smart Library: distributed system for fully autonomous, delocalised book lending across automated totems. IoT hardware, the architecture around it, and the loan logic.
• Best Client: Android app that surfaces a customer’s rating while the phone is ringing, so restaurateurs know who is calling. Published on the Play Store.
• Colors Zone: real-time COVID-19 restriction zone for any Italian region, with the rules attached to each colour.
• Actor Detector: automatic detection and recognition of the actors present in a movie scene.
• ML Algorithms: classification, regression, clustering, dimensionality reduction and ensembles, plus model selection and evaluation.