Hi, I'm Petrit.

Model: claude-sonnet-4-6 — cheaper per token, good enough for the classification loop that dominates an audit.

Pipeline: Bash queue runner with sentinel recovery, AUP refusal detection, usage-limit handling. Per task a fresh CLI process; status file as success interface.

Verification: every promising hypothesis goes to a separate sub-agent with the brief "falsify". Inversion is half the work.

Reports: H1-formatted, with source refs, CVSS rationale, cleanup block. Disclosure runs coordinated with the vendor.

→ Details in the writeups CVSS 9.6 and Catching refusals.

Model: claude-opus-4-7.

What Pixel does: pushes feature branches directly into the puzzlekreis repo, no commenting on tickets. When a run finishes, the result is a feat/... or fix/... branch with tests, not a diff in a chat reply.

Contributions so far: 17 analytics hook sites, photo-gallery lightbox, brand migration, FTS migration with a Postgres array-IMMUTABLE wrapper, Sentry CSP config, a series of pre-soft-launch bugfixes.

Conventions: reads CLAUDE.md per subdir before writing, never git push --force, persistent token telemetry in usage-tracker.json.

→ More in the writeup Pixel — an LLM co-maintainer in production.

Model: claude-opus-4-7.

Job: everything between "keep the home network running" and "write up X for me". Maintains an INFRA.md with hosts, services, domains, accounts, risks, and agents as a living ops wiki.

Heartbeat: every 30 minutes — short status check, on-demand reachable via Telegram.

This site was first-drafted by Nova with an editorial pass from me.

What: Python daemon on the zoopa VPS, running since early March.

Sources: Deutsche Bahn IRIS-TTS (plan + fchg every minute), Open-Meteo weather (every 10 min).

Stations: 11 — five in NRW, two in the Ruhr area, four major hubs (Frankfurt, Berlin, Hamburg, München).

Storage: SQLite WAL, ~425 MB, growing by ~8,000 stops per day.

Phases: 1 collection (live), 2 EDA (done), 3 delay model (May), 4 live dashboard and prediction API.

→ Details in the writeup 384,000 trains, one SQLite.

Hardware: MSI Cubi 5 PRO ADL-U · Intel Core i3-1215U (6 cores / 8 threads, hybrid 2P+4E, 4.4 GHz turbo) · 16 GB DDR5 SODIMM · 256 GB NVMe (Phison) internal · 1 TB external HDD and 256 GB external NVMe for backups.

OS: Ubuntu 24.04, kernel 6.8. ~10 W idle.

Network: behind a Fritz!Box, no port forwarding. Remote access exclusively through Tailscale.

What's running:

• 4 Claude agents (Recon, Pixel, Nova, Atelier) with their own workspaces and Telegram bots
• paradebox — a WiFi-to-Bluetooth audio bridge to my Edifier speaker (Spotify Connect, AirPlay, BlueZ)
• uptime-kuma for monitoring
• redroid container for mobile-app testing

Backups: restic encrypted to the external HDD, weekly. Plus a nightly pull-sync of the zoopa VPS and the puzzlekreis VPS onto paradecentral as an off-site copy.

Hardware: AWS EC2 t3.small in us-east-1 · 2 vCPU Intel Xeon · 2 GB RAM · 30 GB gp3 EBS.

Auth: IAM role on the instance instead of static access keys; permissions delivered through IMDSv2.

OS: Ubuntu 24.04, everything as systemd units, nginx as reverse proxy, Letsencrypt cert with auto-renew.

What's running:

• Main agent (claude-opus-4-7) for VPS maintenance, Discord channel, cross-agent help
• LiteLLM proxy to eight providers: Anthropic, OpenAI, Google, Mistral, DeepSeek, xAI, Cerebras, Perplexity
• several public services under api.zoopa.org/... (civic-tech and side projects) — see zoopa.org
• this site (petrit.dev)

Hardware: Hetzner CX33 · 2 vCPU · 8 GB RAM · 80 GB SSD.

Backend: FastAPI + async SQLAlchemy + Postgres 16 + Redis 7 + Alembic migrations.

Web: Next.js 14 as a PWA.

Mobile: React Native, built via EAS, shipped to Android and iOS.

Size: ~150 backend files, 30+ pytest suites, ~14k lines of TS/TSX in the frontend.

Co-maintained by Pixel — see writeup. Soft launch in the coming weeks.