All case studies
01 — Own project · in daily use
An agent that works while I sleep
A personal AI agent system driven from Telegram: a daily job-offer digest, notifications from long-running work and custom agent skills.
Problem
Every day I checked several job-offer sites by hand, and information was scattered across email, calendar and notes. My coding agent was only usable at the desk — long-running work had to be watched in a terminal.
What I built
- Telegram ↔ agent CLI bridge — talk to the coding agent from a phone: sessions, device pairing with a one-time code, switching the working directory, and a notification when a long task finishes.
- Daily offer digest triggered by a systemd timer every day at 12:00: the agent goes through 13 offer sources following a written procedure, filters offers against a profile and thresholds, and sends a summary to Telegram, split by section to fit the message length limit.
- Task loops — the agent picks up tasks written in Markdown notes and reports the result.
- About 80 agent skills, including my own for Gmail, Google Calendar, Google Drive and OneDrive, Proxmox, Cloudflare (DNS and Tunnel) and OPNsense, plus an idea registry and a user profile.
- Two instances — in the homelab and on a VPS; when the primary fails, the standby automatically takes over the bot and the scheduled jobs.
- Secrets outside the code — in the system keyring and in owner-only files.
Outcome
- A daily offer report without browsing the sites manually: each run checks 20 to 300 offers, and about 5–10% of them make it into the report.
- An offer register of about 1,270 offers in 10 days.
- Continuity through automatic takeover by the second instance.
- Controlling the agent from a phone, even when the work takes longer than I'm at the desk.
- Secrets kept out of the code and out of the repository.
Facing a similar problem?
Describe your process — I'll tell you whether and how it can be solved.