CareerNorth — AI Job Hunt Toolkit

Job hunting is a system design problem: resume parsing, job discovery, fit scoring, resume tailoring, and application submission form a pipeline where each stage must handle failures gracefully. CareerNorth treats this as a multi-stage pipeline with idempotent operations, so a crash at any stage doesn't lose progress. The dual interface (web + CLI) makes the same engine accessible from different workflows — a developer might use the CLI for batch applications while a job seeker prefers the web dashboard.

The system uses a pipeline architecture with clear stage boundaries: Parse Resume → Score ATS → Match Job → Tailor Resume → Apply. Each stage is independently executable and idempotent. The core engine is shared between the Next.js web app and the CLI tool. Resume parsing uses pdfjs-dist for PDF extraction. ATS scoring analyzes keywords, sections, contact info, action verbs, quantification, bullet structure, and formatting. Job discovery uses Playwright to automate browser sessions across 12 platforms (LinkedIn, Naukri, Indeed, Cutshort, Wellfound, etc.). The LLM integration layer supports OpenRouter, OpenAI, and Anthropic for genuine AI resume rewriting — with a deterministic fallback when no LLM key is provided.

The data layer uses Prisma with PostgreSQL for persistent storage (Users, Resumes, Jobs, Applications, Activity logs) with localStorage as the default lightweight backend. Each resume stores parsed sections, ATS scores per job match, and tailored versions. Applications track platform, status, tailored resume snapshot, and timestamp. The model supports per-account history and batch operation tracking.

The hardest challenge was making browser automation reliable across 12 different job platforms, each with unique DOM structures, CAPTCHA systems, and anti-bot measures. Solved with platform-specific selectors, session persistence, and configurable approval gates. Resume tailoring required balancing ATS optimization with readability — aggressive keyword stuffing improves scores but produces unnatural resumes. The solution uses weighted scoring with configurable thresholds per job fit level.

The pipeline is embarrassingly parallel — each job application is independent and can be batched. The CLI uses PM2 for process management. The web app leverages server-side rendering for the dashboard and API routes for the pipeline. Job discovery results are cached to avoid re-scanning platforms. Application rate limiting per platform prevents account restrictions.

Each pipeline stage has retry logic with exponential backoff. Failed applications are logged with full context (platform, error, input state) for manual review. Browser automation handles CAPTCHAs by pausing for human intervention. Session expiration triggers re-authentication prompts rather than silent failures.

Application history provides per-platform success rates, response rates, and timing. ATS score distributions help optimize resume strategies. Pipeline stage timing identifies bottlenecks. Structured logging with correlation IDs connects resume parsing, scoring, and application outcomes.

Prisma was chosen over raw SQL for type safety and migration management. Playwright was chosen over Puppeteer for better cross-browser support and built-in waiting mechanisms. Shared engine architecture (web + CLI) was chosen over separate implementations to maintain scoring consistency. PostgreSQL over SQLite for production reliability and concurrent access.