JOB APPLICATION AGENT

Context

Job hunting at scale is mostly logistics. Read 200 listings, filter for fit, tailor a CV, write a cover letter, fill the form, log it somewhere, follow up if no reply. Multiply by every platform — LinkedIn, Naukri, Indeed, Welcome to the Jungle, Tier-1 company career pages — and the math stops working. I wanted a pipeline that did all of it, ran every few hours, never missed a job, and cost zero.

The Problem

Most "AI job search" tools are thin wrappers around an LLM with no actual scraping, no scoring discipline, no auto-submit, no tracking. The hard parts — surviving anti-bot defenses, scoring with intent (not just keyword match), tailoring per-job, and not getting flagged — are exactly where the value lives. I wanted to build that hard part.

How I Approached It

Four-stage pipeline with hard floors at every gate:

1. Scrape — five concurrent scrapers, each tuned for its source.
2. Score — a local LLM evaluates fit against a strict profile with a hard ₹10 LPA floor.
3. Tailor — only matches ≥ 70 get a CV variant + a 250-word cover letter.
4. Submit — three executor agents handle different submission patterns. Everything logged.

The architecture is deliberately offline-first: no cloud LLM bills, no API throttling, no surprise charges.

What I Did

• Five scrapers (every 4–6 hours): LinkedIn, Naukri, Indeed, Welcome to the Jungle, and Tier-1 company career pages. Each runs in its own Playwright + Chromium context with its own selectors, rotation, and dedup logic.
• Local LLM scoring with Ollama + Qwen 2.5 7B Instruct: Every scraped job gets scored 0–100 against my profile. The prompt includes hard requirements (₹10 LPA salary floor, location constraints, exclusions). Originally used Anthropic Claude API, migrated to local Ollama for $0/month and full privacy.
• Match gate at 70: Anything below disappears. Anything at or above triggers the tailoring stage.
• Two CV variants: A — Operations and B — AI. The scoring LLM also picks which variant fits the role better.
• 250-word cover letter generator: Each one references three concrete details from the listing — no generic templates.
• Three executor agents:
  • India — auto-submits via Playwright on standard ATS systems
  • Abroad — same pattern, with timezone-aware scheduling
  • Tier-1 / FAANG-style — stops short of submit, hands me a one-click confirmation page (these systems detect bots aggressively)
• Google Sheets logging: Every application — job, company, salary, score, CV variant used, cover letter, date, status, link — flows into a single sheet with timestamps.
• Local HTML dashboard: Counts applications by source, scoring distributions, response rates. Refreshes from the sheet via Google Sheets API.
• 11 n8n workflows orchestrating everything: Triggers, retry queues, dedup, sheet writes, follow-up watchers. The whole flow survives crashes — n8n + service account auth survives restarts cleanly.

The Outcome

A pipeline that runs on a local machine with zero monthly cost. Folder structure, all 11 n8n workflows, three Playwright scripts, both CV variants, the Ollama + Qwen install, and the dashboard are done and pushed to GitHub at github.com/rajputdev77-art/job-application-agent. Notion mirror for documentation.

Pending: enabling autostart on the n8n service, finishing the 15-day no-reply follow-up watcher, and wiring the Google Sheet service-account credentials end-to-end.

Update: Sharper filters and an always-on dashboard

June 2026 — tightening the pipeline now that it was running on real volume:

• Junk-job pre-filter. Before scoring, the pipeline now drops obvious non-fits (HR, sales, manufacturing roles) so the local LLM never wastes cycles scoring them — plus a Naukri-specific ₹10 LPA salary filter at the scrape stage, cutting noise before it reaches the model.
• Persistent dashboard server. The local HTML dashboard now runs as a persistent Node server with Windows environment variables wired in, instead of being launched by hand — so the application stats are always up when I glance at them.
• One-command boot. Boot scripts now start all four services together — n8n, the helper, the dashboard, and Ollama — so the whole stack comes up with a single action instead of four.

What I Learned

• The migration from Claude API to local Ollama was the single biggest unlock. Free, private, faster on most scoring tasks. The pipeline became sustainable overnight.
• Selector brittleness is the real cost of scraping. Half the engineering work is keeping selectors alive across UI updates. Playwright's auto-wait + retry logic absorbed most of it.
• Three executor agents is not over-engineering. ATS submission patterns vary enormously between regions and company tiers. A single agent would have failed at ~30% of submissions.
• The hardest part of "AI job search" is not the AI. It's the integrations, the rate limits, the sheet writes, the dedup, the failure recovery. The intelligence is a one-line API call. Everything else is the system.