Job Application Agent

The system that may have brought you here. An autonomous agent that discovers relevant job postings, scores them against my background, generates tailored application materials in my voice, and sends everything to my phone for approval before anything goes out.

Python qwen3-next-80b SQLite Telegram Gmail API
Pipeline demo coming soon

How It Works

The agent runs as a persistent daemon on my Strix Halo node, integrated into the broader agent network. It follows a six-stage pipeline from discovery to interview.

Pipeline

  • Discovery — Scrapes HN "Who is Hiring" threads via Algolia API, RSS feeds from AI job boards, and manual URL submissions via Telegram
  • Scoring — Each posting is evaluated by qwen3-next-80b against my background (game dev, AI tools, mocap, agent systems). Scores 0.0–1.0 with reasoning.
  • Notification — High-relevance matches (>60%) are pushed to my phone via Telegram with one-tap approve/skip
  • Tailoring — On approval, the agent generates a resume and cover letter customized for the specific role, using a voice profile built from my actual writing
  • Review — Every outbound document goes through human review. I see the full draft in Telegram before anything sends.
  • Coordination — Tracks application status, recruiter responses, and interview scheduling in a SQLite pipeline database

Why Build This?

Job searching is high-volume, repetitive, and rewards customization — a perfect target for an AI agent. But it also requires judgment, voice, and relationship management that AI can't fully automate. The human-in-the-loop design means I get the efficiency of automation with the quality control of personal attention.

It's also a portfolio piece. If you're reading this because you're hiring for AI tooling roles, this system is literally the job description in action: local LLM inference, agent orchestration, human-in-the-loop design, and production-grade Python infrastructure.

Integration

The agent uses every part of the network: inference lock for model coordination, dispatch routing for LLM selection, the Telegram bridge for notifications, semantic memory for learning from past applications, and fleet coordination for visibility. It's not a standalone script — it's a citizen of the agent network.