Job Hunter — Ireland 🇮🇪

An AI-powered job search assistant I built specifically for the Irish job market.

Why I built this

While job hunting in Ireland, I ran into a real problem: entry-level, graduate, intern, and apprenticeship opportunities are scarce and easily lost in the noise. The big job boards bury them under senior roles, recruiter spam, and non-software listings, and no single board has good Ireland coverage on its own.

So I built my own pipeline. Based on my own profile (MSc Computer Science @ UCD, Java / Python / Spring Boot / AI-ML / Full Stack), it:

1. Scrapes jobs across every major source — LinkedIn, Indeed, Glassdoor, IrishJobs, Reed (UK/IE), plus direct company career sites (Workday, Greenhouse, SmartRecruiters, Lever, Workable) — all filtered down to genuinely entry-level Irish roles.
2. Scores every job out of 10 using a strong, profile-aware AI prompt — with a verdict (Strong Apply / Apply / Maybe / Skip) and a breakdown.
3. Analyses the gaps between my saved resume and each job description.
4. Tailors my resume into an ATS-optimised one-page DOCX, applying only the fixes I accept.
5. Writes a tailored cover letter for the role.

My resume lives in the DB/profile once, so every job gets scored and tailored against it automatically.

---

Features

Job aggregation (free, no paid scraping)

• LinkedIn — public guest jobs API (no login), real Ireland listings with posting dates
• Indeed · Glassdoor · IrishJobs — via the Jooble aggregator API (free key), labelled by their origin board
• Reed — UK/Ireland job board API
• Workday — direct CXS API for 27 verified Ireland-office employers (Mastercard, Cisco, State Street, Stryker, Medtronic, Accenture, Intel, AstraZeneca, Pfizer…)
• Greenhouse — public ATS API for ~78 Dublin-office tech companies (Stripe, Intercom, Cloudflare, MongoDB…)
• SmartRecruiters · Lever · Workable — more direct company career sites
• BambooHR is supported but needs a per-company list, so it's low-yield and off by default.

Smart filtering (the core of it)

• Entry-level only — a word-boundary regex rejects senior / staff / principal / lead / director / manager / architect / Engineer II-III etc., while keeping genuine junior / graduate / intern / associate roles
• Experience filter — rejects descriptions requiring "3+ years", "5+ years" ranges, or intern-exclusion clauses
• Software-only — drops sales, marketing, HR, finance, physical/hardware engineering, pharma/medtech, and support roles
• Filters run twice: at scrape time (before saving) and as a post-search DB purge, so junk never re-appears

AI scoring & tailoring

• Score out of 10 with verdict, one-line reason, and a breakdown (required skills / preferred skills / cultural fit / ATS keyword coverage)
• Matches, gaps, and red flags per job
• ATS keyword analysis — which keywords from the JD are present vs missing in my profile
• Gap analysis — honest, profile-grounded suggestions (each tagged Safe or Stretch) for closing the gap
• Resume tailoring — generates a one-page ATS-optimised DOCX applying only the gap fixes I accept
• Cover letter — tailored, human-sounding letter for the specific role
• Paste-any-JD — paste a description from anywhere to score + tailor against it instantly

Workflow UI (dark theme)

• Jobs sorted past-24h first, then newest, with the posting date (not scrape date) shown on each card
• Source badges, score, and verdict on every card
• Applied / Delete buttons that hide a job and stop it reappearing on future scrapes (status is preserved across re-scrapes)
• ↺ Restore to undo a mis-clicked Applied/Delete
• Live keyword filter over loaded jobs + score/status/source filters

---

How I did it

Architecture

┌────────────────┐     ┌──────────────────────┐     ┌─────────────────┐
│  Next.js (TS)  │────▶│   FastAPI backend    │────▶│  Supabase (PG)  │
│  dark-theme UI │◀────│  scrape · score · AI │◀────│  jobs table     │
└────────────────┘     └──────────┬───────────┘     └─────────────────┘
                                   │
            ┌──────────────────────┼───────────────────────────┐
            ▼                      ▼                            ▼
   Job sources (free)      Scoring engine              Tailoring (Claude)
   LinkedIn · Jooble       Ollama (local, free)        gap analysis
   Reed · Workday          → Claude Haiku fallback      resume DOCX (Node/docx)
   Greenhouse · etc.                                    cover letter

The pieces

• backend/scraper.py — one scrape_*() per source, all normalised to a common job shape, run in parallel/sequential as each API tolerates. _is_engineering_role() is the three-layer entry-level + software-only filter applied before anything is saved.
• backend/profile.py — my candidate profile, resume text, target roles, search queries, and the exact scoring/resume/cover-letter prompts.
• backend/scorer.py — scores each job 0–10. Uses Ollama locally for free when available, and falls back to Claude Haiku otherwise. Returns verdict, reasons, matches/gaps, keywords, and a weighted breakdown.
• backend/tailor.py — Claude (Sonnet) for gap analysis, resume JSON, and cover letters. The resume JSON is rendered to a real .docx by a small Node.js docx generator (backend/docx_gen/generate.js).
• backend/db.py — Supabase CRUD. Upserts dedupe by URL and preserve user-set status (applied/rejected) so hidden jobs never resurface. purge_bad_jobs() re-applies the filters after each search.
• backend/main.py — FastAPI endpoints (search, jobs, status, scoring, tailoring).
• frontend/app/page.tsx — the entire single-page dark-theme UI.

Design decisions

• Free-first: every source is a free/public API; scoring prefers local Ollama; Claude is used only where quality matters (tailoring). Jooble's free key is used sparingly (≈6 requests per full search).
• Filter twice: bad jobs are filtered both before insert and after search, because new sources keep finding new ways to sneak senior/non-software roles in.
• Posting date, not scrape date: each source's real posted date is parsed and normalised to ISO so "past 24h" sorting is meaningful.

---

How to run

Quick start (Windows)

After one-time setup below, just double-click start.bat in the project root. It launches the backend (port 8000) and frontend (port 3000) in their own windows — safe to re-run, it skips anything already running — and opens the app in your browser. If the UI ever shows "Backend offline", re-running start.bat fixes it.

Also add APIFY_TOKEN=... to backend/.env (optional) to unlock real Indeed + Glassdoor Ireland listings via Apify actors (~$0.03 per full search, free $5/month credit).

Prerequisites

• Python 3.11+
• Node.js 18+ (for the frontend and the DOCX generator)
• A free Supabase project (Postgres)
• An Anthropic API key (for tailoring; scoring can run free on Ollama)
• Optional but recommended: Ollama running locally with llama3.2 for free scoring
• Optional: free API keys for Reed and Jooble

1. Database

Create the jobs table — run the SQL in SETUP.md (or SCHEMA_SQL in backend/db.py) in the Supabase SQL Editor.

2. Backend

cd backend
pip install -r requirements.txt

# Node deps for the DOCX generator
cd docx_gen && npm install && cd ..

# Create backend/.env (this file is gitignored — never commit it)

backend/.env:

ANTHROPIC_API_KEY=sk-ant-...
SUPABASE_URL=https://xxxx.supabase.co
SUPABASE_ANON_KEY=...
REED_API_KEY=...          # optional
JOOBLE_API_KEY=...        # optional (free, unlocks Indeed/IrishJobs/Glassdoor)
ADZUNA_APP_ID=...         # optional
ADZUNA_API_KEY=...        # optional
RAPIDAPI_KEY=...          # optional

Run it:

uvicorn main:app --port 8000

(Optional) free local scoring:

ollama pull llama3.2
ollama serve

3. Frontend

cd frontend
npm install
npm run dev

Open http://localhost:3000.

4. Use it

1. Click Full Search in the header — it scrapes every source, filters to entry-level Irish roles, and scores them (browse as they populate).
2. Click a job to see its score breakdown, matches, gaps, and ATS keywords.
3. Hit Analyse Gaps → accept the fixes you want → Tailor Resume to download an ATS DOCX.
4. Hit Cover Letter for a tailored letter.
5. Use ✓ Applied / ✕ Delete to keep your list clean (they won't reappear on re-scrapes); ↺ Restore to undo.
6. Or paste any JD into the sidebar box to score + tailor against it instantly.

---

Tech stack

Layer	Tech
Frontend	Next.js (App Router) · React · TypeScript · Tailwind CSS
Backend	FastAPI · httpx · Pydantic
Database	Supabase (PostgreSQL)
Scoring	Ollama (local, free) → Claude Haiku fallback
Tailoring	Claude Sonnet + Node.js docx generator
Sources	LinkedIn · Jooble · Reed · Workday · Greenhouse · SmartRecruiters · Lever · Workable

---

Notes & honesty

• LinkedIn/Indeed/Glassdoor/IrishJobs block direct free scraping. I get them legitimately via LinkedIn's public guest API and the Jooble aggregator. There's no paid scraping in the default flow.
• This is built around my profile and resume. To use it for yourself, edit backend/profile.py (profile, resume text, target roles, search queries).
• backend/.env holds secrets and is gitignored — never commit it.

Built by Geet Bhute · Dublin, Ireland.