goosehunt

A personal tool for UWaterloo co-op students that turns WaterlooWorks job boards (Employer Direct and Full Cycle Service) into a ranked, filterable local UI — scored against your resume and classified by role.

Personal use only. WaterlooWorks ToS likely prohibits automated scraping. Use at your own risk.

---

What it does

1. Scrapes your currently visible WaterlooWorks results using Playwright (Employer Direct or Full Cycle). You log in manually, set your filters and work term, then press Enter — the scraper does the rest.
2. Stores every posting in a local SQLite database. Resumable: a crashed scrape picks up where it left off.
3. Classifies each posting against four role types (SWE, AI/ML, firmware, hardware) using tunable keyword lists in config/roles.yaml.
4. Scores each posting against your resume PDF using cosine similarity on sentence embeddings.
5. Serves a local web UI — one page, all postings loaded, client-side sort/filter, keyboard navigation, no build step.

---

Requirements

• uv — Python package manager
• A UWaterloo WaterlooWorks account with active co-op access
• Your resume as resume.pdf

---

Setup

make install
cp /path/to/resume.pdf resume.pdf

make install creates a venv, installs dependencies, and downloads the Chromium browser for Playwright.

---

First run

make run                  # Employer Direct (default)
make run BOARD=full_cycle # Full Cycle Service

This opens a Chromium window. Log in to WaterlooWorks (Duo if prompted), navigate to the board you chose, set your work term and filters, wait for job listings to appear, then press Enter in the terminal. The scraper runs, the pipeline processes everything, and the UI starts at http://localhost:8000.

On subsequent runs where you just want to re-serve existing data:

make serve

To re-scrape and reprocess without restarting the server:

make scrape BOARD=full_cycle && make pipeline

---

Docker (any device, no Python setup)

Scraping must still run locally (it needs your WaterlooWorks session). Everything after that runs in the container.

# scrape locally first
make scrape

# on any device with Docker — copy the repo + data/postings.jsonl + resume.pdf, then:
docker compose up

docker compose up runs ingest → embed → score → serve on every start, then keeps the UI alive at http://localhost:8000. The HuggingFace model is cached in a named volume so it's only downloaded once.

---

Makefile targets

make install     # create venv, install deps, install Chromium
make run         # scrape + pipeline + serve (BOARD=direct by default)
make scrape      # scrape only → data/postings.jsonl (BOARD=direct|full_cycle)
make pipeline    # ingest → embed → score (run after scrape)
make serve       # start FastAPI on localhost:8000
make test        # run unit tests (no browser required)

Set BOARD=full_cycle on make run or make scrape for the Full Cycle board.

Individual pipeline steps: make ingest, make embed, make score.

---

UI

Two-pane layout: sortable table on the left, posting detail panel on the right.

Full UI — sortable table with resume, role, and pay scores across all postings

Detail panel — score grid, matched keywords per role, and apply link for a selected posting

Sort by any score — here sorted by hardware relevance

Search — instant client-side filter across title, org, summary, and skills

Command palette (Ctrl+K) — keyboard-driven access to all sort and filter actions

Filters: search box, role chips (SWE, AI/ML, FW, HW), apply-by chips (Email, Link).

Table columns: title, org, location, deadline, resume score, role scores, pay, openings, status. All sortable. Click a job ID to copy it.

Detail panel: score grid, apply link/email with copy buttons, keyword-hit chips showing which keywords fired per role, summary/responsibilities/required skills, and local status buttons (New / Maybe / Applied / Ignored).

Keyboard shortcuts:

Key	Action
j / k	Navigate rows
/	Focus search
Esc	Blur search
c	Copy job ID
m	Copy apply email
Shift+S	Sort by resume score
Shift+P	Sort by pay
Ctrl+K	Command palette

A Day/Night toggle persists the theme in localStorage.

---

Scores

Score	What it measures
score_software	SWE keyword match
score_ai_ml	AI / ML / data science keyword match
score_firmware	Firmware / embedded / mechatronics keyword match
score_hardware	Hardware / FPGA / PCB keyword match
score_resume	Cosine similarity between posting and your resume
comp_score	Estimated hourly pay, normalized to [0, 1] ($16/hr → 0.0, $60/hr → 1.0)

All scores are in [0, 1]. Keywords are tunable — edit config/roles.yaml and rerun make score.

---

Customizing roles

Edit config/roles.yaml to add or remove keywords for any role, then rerun scoring:

software:
  keywords:
    - Python
    - Kubernetes
    - distributed systems
    - quant        # add anything relevant to your background

make score

This re-scores all postings in under a second. No re-scraping or re-embedding needed.

---

Refreshing postings

make scrape is resumable and skips job IDs already in data/postings.jsonl. To force a full refresh after changing the scraper or parser:

mv data/postings.jsonl data/postings.old.jsonl
make scrape
make ingest
.venv/bin/python embed/embed_postings.py --force
make score

Local job statuses (new/maybe/applied/ignored) are stored in data/postings.db and are preserved across re-scrapes.

---

Project layout

goosehunt/
├── config/
│   └── roles.yaml          # keyword lists for each role scorer
├── scraper/
│   ├── scraper.py          # Playwright scraper → JSONL
│   └── test_scraper.py     # unit tests (no browser required)
├── db/
│   ├── schema.sql          # CREATE TABLE statements
│   └── ingest.py           # JSONL → SQLite
├── embed/
│   ├── embed_postings.py   # sentence-transformers → BLOB column
│   └── embed_resume.py     # embed resume PDF → score_resume column
├── classifier/
│   └── scorer.py           # keyword scorer → score_* columns
├── resume/
│   └── parser.py           # pdfplumber PDF → plain text
├── web/
│   ├── main.py             # FastAPI app
│   └── static/
│       └── index.html      # Alpine.js UI, no build step
├── scripts/
│   └── preflight.py        # input validation before pipeline/Docker
├── data/                   # gitignored — created at runtime
│   ├── postings.jsonl
│   └── postings.db
├── resume.pdf              # gitignored — add your own
├── Dockerfile
├── docker-compose.yml
└── Makefile

For implementation details and design decisions, see DESIGN.md.

---

License

MIT

---

Notice

goosehunt is not affiliated with the University of Waterloo, WaterlooWorks, or Orbis. It does not bypass login, Duo, or any WaterlooWorks access controls — the scraper runs through your own browser session after you manually log in.

All scraped data is stored locally. Nothing is sent to a goosehunt server (there isn't one). Users are responsible for complying with WaterlooWorks terms, university policies, and any applicable laws before scraping or storing posting data.

goosehunt is provided as-is, without warranty. Verify important details in WaterlooWorks or with the employer before making application decisions.