Blog · RSS

Notes on what I build

I document my work in public: technical decisions, mistakes and what I learn along the way.

• October 10, 2026 · 5 min readFrom 13 GB to 2.7 GB: choosing format and resolution for scanned mangaA series of scanned manga weighed 800 MB per volume. I compared JPEG, WebP and AVIF on real pages, found the point where the screentone breaks down and brought it to a fifth of the size.#homelab#python#performance
• October 10, 2026 · 4 min readtranscribe: from one voice note to an open-source terminal toolI needed to turn a voice note into text without uploading it anywhere. That same night the ten-line script became an interactive CLI with subtitles, tests and an MIT license.#python#ai#open-source#cli
• October 7, 2026 · 6 min readNotebook: my notes as a chat with myself, now open sourceIt started as a two-day experiment so I would stop losing my university notes. Two months later I rebuilt it measuring everything, kept it running for free and released the code under the MIT license.#ai#rag#open-source#node
• September 26, 2026 · 7 min readYour own Netflix at home: a media center with Jellyfin, Docker and secure accessHow I built Selflix, my Jellyfin media center on a modest home server: Docker, the folder structure Jellyfin expects, access from outside without opening ports, per-person accounts and the pitfalls that cost me hours.#homelab#linux#docker
• September 24, 2026 · 5 min readLong-running LLM jobs: multiple providers, a resumable queue and live progressAnalyzing a contract clause by clause can take minutes. In AdvAI I solved it with a model pool that respects each provider's limits, a worker that survives restarts and Server-Sent Events to show progress.#ai#llm#python#backend
• September 23, 2026 · 5 min readAn AI agent that runs my server: memory in files and skills as proceduresMy homelab is maintained by an agent (OpenClaw) running 24/7. What makes it reliable isn't the model: it's memory written in Markdown, explicit red lines and procedures with success criteria. Here's how I organized it.#ai#linux
• September 22, 2026 · 5 min readAnonymizing a contract before sending it to an LLM (without another LLM)In AdvAI no name, ID number or tax ID leaves the machine: a rule-based anonymizer swaps them for stable placeholders and the system restores them in the answer. How it works and why I didn't use AI for it.#ai#llm#privacy#python
• September 21, 2026 · 6 min readThesis PDFs: extracting data with AI without trusting it with everything, and showing them without eating all the RAMIn SGPG, my university's thesis management system, every PDF is processed in the browser: deterministic rules where possible, an LLM where it pays off and heuristic fallbacks. Plus a viewer that only draws the visible pages.#ai#llm#nextjs#performance
• September 19, 2026 · 5 min readSwitching bootloaders without fear: test Limine once with BootNextFor rOS, my flavor of Arch, I replaced GRUB with Limine without risking the boot: install it side by side, test it on a single reboot and promote it only if it works. Idempotent scripts, pacman hooks and atomic copies to the ESP.#linux
• September 18, 2026 · 5 min readTruly offline-first: a water billing app with Electron and SQLiteLink'u records meter readings and issues bills for a rural community where the internet isn't guaranteed. SQLite as the source of truth, all logic in the main process, rates frozen on every bill and AES-256-GCM encrypted backups.#electron#react#security
• September 17, 2026 · 5 min readA citizen report map with react-leaflet (and how I'd rewrite it today)At the Hackacom 2023 hackathon we built Canasta: a map where neighbors report problems in the city and others join in to fix them. I go over the real code, the classic Leaflet icon bug and what I'd change three years later.#react#frontend
• September 16, 2026 · 4 min readCoding before and after AI: what years of doing it by hand left meI learned the old-school way, and we only brought AI into our work during my last year at WANT. What changed, what didn't, and how I decide when to let AI move fast and when to review with a magnifying glass.#ai#career
• September 15, 2026 · 4 min readHow I stopped an LLM from inventing articles of the Civil CodeIn AdvAI, my capstone project, the model doesn't write legal citations: it only picks from real sources and the system inserts the verbatim text. That's how I went from "sounds right" to verifiable citations.#ai#llm#rag#python
• September 13, 2026 · 6 min readHow I built this portfolio: one file for the website, the resume and two languagesThis site and my PDF resume come from the same data file. Astro, GSAP animations driven by attributes, a WebGL shader with a CSS plan B, a resume that passes ATS filters and a subdomain solved with 10 lines at the edge.#frontend#performance