AI / ML system
Hermes Agent — Job Discovery Pipeline
AI Automation · Solo · Self-hosted on Hostinger VPS
Autonomous AI agent on a Linux VPS: scans job & freelance platforms daily, filters and ranks listings against a skill profile, and delivers a curated Telegram digest every morning at 9 AM — fully unattended.
Model pipeline
Research → result
Lab-style panels for training, inference, and outputs.
System design
Overview
A production-grade autonomous agent built on the open-source Hermes Agent framework. It runs a scheduler-driven pipeline on a Hostinger VPS — no manual babysitting — that discovers opportunities, applies LLM-powered filtering and ranking, and pushes only the best matches to Telegram before you start your day.
What I built
- Hermes Agent framework — autonomous scan → reason → act → observe loop
- Multi-source ingestion: job boards, freelance platforms, and listing feeds
- LLM-powered filter, rank, and summarize against a configurable skill profile
- Telegram Bot API delivery — formatted morning digest at 9:00 AM
- Hostinger VPS deployment with systemd unit + timer (persistent, restart-on-failure)
- Provider/model integration with fallback handling for entitlement and quota errors
- Framework-level patches for streaming response crashes in production SDK paths
- Environment-isolated secrets (.env / EnvironmentFile) — no keys in repo
What's unique
- End-to-end ownership: architecture, agent reasoning loop, LLM integration, and Linux deployment
- Unattended systemd service — daily cron at 9 AM with persistent timers and restart policies
- Production debugging at framework source level: patched streaming SDK crashes and provider entitlement failures
- Skill-profile matching — noise filtered out before anything hits your inbox
Why it matters
- Real AI automation that runs while you sleep — not a demo script
- Shows ops maturity: VPS, systemd, Telegram delivery, and LLM provider resilience
- Recruiters see agent design + production incident response, not just prompt engineering
Measured gains
Results that matter
Same model family — gains come from forensic-grade data engineering, not architecture tricks.
9 AM daily
Delivery
Telegram digest · fully scheduled
Unattended
Runtime
systemd service on Hostinger VPS
Scan→Rank→Send
Pipeline
Autonomous Hermes agent loop
Production
Deploy
Self-hosted · not a local notebook
Figures & graphs
See the pipeline clearly
Architecture, selective Real-ESRGAN, YOLO detections, training curves, and mobile/web deployments.
Craft
How the intelligence was built
Open the research