I build software products, and mods on the side.

Law firms, courts and institutions have to strip names, ID numbers, addresses, faces and signatures out of documents before sharing them, and most tools only draw a black box on top. Censory finds the personal data with a vision model, then removes it from the PDF content stream itself, with a person approving every redaction.

Files are encrypted at rest and never written to disk while detection runs. My recent work: the redaction editor rebuilt as a workbench, a touch shell for phones, a benchmark of vision models paired with deterministic face detection, and the pricing and marketing pages in English and Romanian.

AI-written blogs that live on the customer's own domain instead of a subdomain nobody links to. A Go API and worker generate, review and schedule the posts, a SvelteKit dashboard runs the show, and delivery happens through a JavaScript SDK, an edge proxy, or a WordPress or Shopify plugin.

The articles render through MD2, my directive layer over Markdown, and a large share of the work went into making generated text stop reading like generated text. The WordPress and Shopify integrations are open source.

A meal planner for a household that actually cooks. Instead of generating a week of unrelated recipes, it composes meals from a library of building blocks with known macros, respects each person's calorie target, rotates breakfasts, and reuses batch-cooked components across the week.

Nutrition data comes from the USDA FoodData Central exports. Most of the planner logic was written and verified through unattended Claude Code runs, with every decision logged for me to check the next morning.

I work across five machines and three identities, and none of them knows what the others are doing. Jarvis is one always-on hub with a single memory and a Telegram thread as the conversation, plus node agents on each machine that do the work where the work lives, including starting Claude Code in its own worktree and handing back a pull request.

The design was stress-tested before any code existed: a decision log past ninety entries, a vault-based memory model I can open in Obsidian, and hard boundaries around employer data. The hub, nodes, plugin system and dashboard exist today; the user docs are written in MD2.

Reading printed Cyrillic text from scanned pages. A preprocessing module cleans and segments the scans with OpenCV, a feature extraction step turns glyphs into vectors, a Keras network classifies the characters, and a React front end drives the whole thing.

I worked on preprocessing. It is old code and reads like it, but it was the first time I shipped machine learning as part of a team, and the repositories are public.