Applyr Job Pipeline

Applyr is an automated job application pipeline combining multi-platform job scraping, career-level-aware scoring, AI vision-driven form filling, and self-recovery. It runs as a FastAPI service on callisto (port 8585) with an admin UI integrated into Arcturus-Prime at /admin/jobs.

Components

JobSpy Advanced Lab

Multi-platform job scraper with intelligent scoring.

Source: ~/Development/JobSpyAdvanced-Lab/

Capabilities:

Scrapes LinkedIn, Indeed, Glassdoor, Google Jobs, ZipRecruiter
Proxy rotation for reliable access (Sentinel VPS Squid proxy)
Career-level-aware scoring engine
Security clearance detection and filtering
CSV output with detailed score breakdowns

Applyr Service

FastAPI service for automated job applications with AI form filling.

Source: ~/Development/applyr/ Service: rc-service applyr start (OpenRC, port 8585) Git: https://git.Arcturus-Prime.com/KeyArgo/applyr.git

Key Capabilities:

LinkedIn Easy Apply + Indeed Apply automation
AI vision-driven form filling (Claude vision API)
Candidate profile system with verified screening answers
Watchdog, circuit breaker, and per-job timeout self-recovery
Anti-bot detection hardening (Patchright, Bezier mouse, human typing)
AI-generated tailored resumes and cover letters per job
Application pipeline tracking with follow-up scheduling

Key Endpoints:

POST /api/jobs/start — Start apply session (csv, threshold, platforms, profile)
POST /api/jobs/stop — Graceful stop
POST /api/jobs/panic-stop — Emergency browser kill
GET /api/jobs/status — Session status
GET /api/jobs/profiles — List candidate profiles
GET /api/jobs/health — Health check

Arcturus-Prime Admin UI

Dashboard at /admin/jobs for managing job searches and reviewing results.

Proxy Config: astro.config.mjs proxies /api/jobs/* → 10.42.0.100:8585

AI Form Filling

Vision-Driven (Primary)

When enabled (APPLYR_VISION_ENABLED=true), each form step is:

Screenshotted
Sent to Claude vision API with candidate profile + job context
AI returns structured JSON (field type, label, answer, confidence)
Playwright fills elements, matching vision labels to DOM via fuzzy matching
Falls back to legacy selectors for low-confidence fields

Module: applyr/vision.py — VisionFormAnalyzer, FieldAction, VisionFormResult Model: Configurable via APPLYR_VISION_MODEL (default: claude-haiku-4-5-20251001)

Legacy Selector-Based (Fallback)

CSS selector pattern matching with AI text fallback:

Field hierarchy: select → radio (fieldset) → radio (bare) → text → number → textarea → checkbox
Screening KB keyword matching → cache → Claude Haiku fallback
Bare radio handler for employer custom questions (agreement/consent detection)

Candidate Profile System

JSON profiles in data/profiles/ contain identity, career, education, certifications, skills, work history, projects, and screening answers. The active profile overrides the default screening KB during sessions.

Active profile: daniel-laforce.json (61 screening answers, verified from master resume) API: GET /profiles (list), GET /profiles/{slug} (detail) UI: Profile selector dropdown in Start Session modal

Self-Recovery

System	Behavior
Per-job timeout	`asyncio.wait_for()` at 180s — cancels stuck jobs
Watchdog	Background task monitors heartbeats every 10s, alerts at 90s stall
Circuit breaker	Pauses 120s after 5 consecutive failures
Recovery	Escape → dismiss modal → navigate neutral page → recreate page if stuck
Stop signal	`ctx.check_stop()` inside form loops — responsive within 2-3s

Scoring System

Role Categories

Jobs are categorized by title and scored based on fit:

Category	Base Score	Dan’s Level
Core IT (SysAdmin, NetEng, SysEng)	120	Mid-career (15+ years)
Target Engineering (DevOps, SRE, Platform)	100	Entry-level pivot
Entry Level (NOC, IT Specialist)	60	Below level, foot in door
Adjacent Tech (SWE, Data Eng)	20	No professional experience
Remedial (Help Desk)	10	Overqualified, beats Starbucks

Seniority Modifiers

Entry-level DevOps/SRE: +50 boost (total: 150) — #1 target
Entry-level + strong keywords: Additional +30 bonus
Senior DevOps/SRE: -60 penalty (can’t compete)
Senior Core IT: -15 (Dan could do this)

Keyword Scoring

Weighted by relevance to Dan’s actual skills:

High value (8-10): Gentoo, K3s, Tailscale, WireGuard, Proxmox, ArgoCD
Good match (5-7): Kubernetes, Linux, Ansible, Terraform, CI/CD, Python
Negative: Printer repair, call center, nursing, accounting

Environment Factors

Remote: +20, Hybrid: +5
Open source: +15, Linux: +5
100% on-site: -50, Heavy travel: -100
Security clearance required: -1000

Data Files

File	Purpose
`data/profiles/*.json`	Candidate profiles (identity, career, screening answers)
`data/candidate_context.md`	Career positioning guide for AI resume tailoring
`data/screening_answers.json`	Screening answer KB (62 categories, standalone fallback)
`data/resumes/master/`	Master resumes + RAG source
`data/resumes/role_specific/`	9 pre-tailored resume variants
`data/scored_jobs/`	Scored job results by date
`data/session_logs/`	Per-session log files

Quick Reference

# Start the service
sudo rc-service applyr start

# Check status
curl http://localhost:8585/api/jobs/health

# Run a job search
cd ~/Development/JobSpyAdvanced-Lab
./venv/bin/python3 run_jobspy_and_scorer.py