procyon/README.md

Procyon — Job Hunt Workflow Engine

Multi-step job-hunt pipeline with a web GUI, running on a Proxmox LXC (never on the MacBook).

Pipeline

1. Discover — pull jobs from RemoteOK + Indeed RSS (+ optional extra feeds), store in SQLite.
2. Score — fit score (0–1) + a one-line "why" opinion, LLM-first, keyword fallback.
3. Tailor — clone the master resume and reorder/re-emphasize for the role (honest-only: no fabricated employers/dates/skills). Draft a tailored cover email.
4. Self-Audit — automatic checklist (employer name, no placeholders, contact info, subject, resume targets the job). Every item must PASS.
5. Approve (human gate) — you review and click Approve. Nothing sends without this.
6. Send — SMTP email with tailored resume attached. Send is BLOCKED unless approved AND the audit passed AND SMTP is configured.
7. Track — log outcomes (response/interview/offer) and surface response-rate insights.

Design rules

• Inference runs on the LAN Ollama host (llm_base_url), NEVER on this CT (drjones rule).
• Graceful degradation: if the model is saturated, scoring/tailoring falls back to deterministic keyword matching — the pipeline never blocks on inference.
• Honest resumes only. The LLM system prompt forbids fabricating credentials.
• Optional outbound proxy (his own Nord proxy CT) for discovery fetches — not an evasion rig.

Setup

• systemctl restart procyon after deploy.
• Configure SMTP (iCloud app-specific password) in Settings before sending.

Files

• app.py — Flask routes + pipeline orchestration
• db.py — SQLite schema + helpers
• scraper.py — discovery (RSS/JSON + proxy)
• llm.py — Ollama client + keyword fallback
• tailor.py — scoring + resume tailoring + email drafting
• audit.py — pre-send self-audit
• emailer.py — SMTP send
• improve.py — outcome tracking + insights