JobHub

A local internship search engine. It pulls postings from GitHub aggregator lists and company ATS boards, dedupes them, drops the ones that break your hard rules, scores everything that survives, and shows you the shortlist — sorted into Likely, Reach and Wildcard.

Every model call goes through headless Claude Code (claude -p) on your subscription. No API key, no per-token bill.

Most of the work is free. Ingest, dedupe, the prefilter and a deterministic 0–100 score cost nothing. Only the unclear middle — roughly one posting in seven — reaches a model, under a token cap you set. The full rubric never runs on its own; you ask for it per job.

Setup

python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/jobhub init
.venv/bin/jobhub smoke          # checks `claude -p` structured output works

Then fill in profile/profile.md (prose: background, what you have built, what you want — this is the only thing a model reads) and profile/profile.yaml (term, locations, dealbreakers). Both are editable from the web UI's Profile page.

Daily use

jobhub serve                    # http://127.0.0.1:8765
jobhub run                      # ingest -> score -> digest, capped

The web UI is a left rail — Home · Jobs · Applications · Profile · Watchlist · Engine — with Run at its foot.

Home	what arrived since your last visit, what's open, what's in flight, what closes in 14 days
Jobs	bucket tabs, one search box, filters in a drawer; ⚡ deep-evals selected postings
Applications	everything you've acted on, grouped by where it stands
Profile	the constraints that decide what gets scored — editable, with a re-score warning on the ones that matter
Engine	the pipeline stage by stage, what's holding jobs back, live log
Settings	14 themes, scoring thresholds, token budgets

Themes were built in OKLCH and are checked against WCAG AA plus a contrast floor on card edges; the choice is stored in the database, so the CLI and browser agree. The setting survives browsers and reinstalls.

To run it every morning: scripts/install_launchd.sh 7 30 (macOS launchd; logs to data/logs/).

CLI

jobhub ingest [--fetch-limit N]     # read sources, dedupe, fetch descriptions   (free)
jobhub score  [--budget N]          # local score, then capped Haiku triage of the unclear band
jobhub score  --calibrate           # how the local scorer agrees with your deep evaluations (free)
jobhub deep 123 456                 # full rubric on specific jobs; --top 20 --bucket likely
jobhub digest                       # writes digests/YYYY-MM-DD.md
jobhub show --bucket likely         # or reach / wildcard; `jobhub show <id>` for one
jobhub status <id> applied          # new | shortlisted | applied | interview | rejected | offer | skipped
jobhub rescore --carry-over         # after a change that only *widens* acceptance
jobhub rescore --reprefilter        # apply a tightened deterministic rule to jobs already scored (--dry-run first)
jobhub discover                     # Claude researches companies matching your profile
jobhub companies resolve            # find ATS boards for companies missing one

How companies get watched

1. Aggregator lists — the broadest source, read every run. Ingest also searches GitHub weekly for new ones.
2. Harvest — any company whose posting URL reveals a Greenhouse/Lever/Ashby board is added automatically and polled from then on. This is the main breadth lever, and it costs nothing.
3. Seeds in profile/companies.yaml, or the Watchlist page.
4. Discovery — jobhub discover researches companies via web search. Those with a readable board are approved.

Breadth is deliberately cheap: a board with no matching roles is one HTTP call per run, and the filtering happens downstream. Supported boards are Greenhouse, Lever, Ashby, Workday (workday:tenant/wd/site) and SmartRecruiters. Known gap: Oracle HCM pages (JPMC, Amex) are JS-rendered and can't be fetched, so those postings stay text-less and are scored on their title alone rather than being dropped.

Tuning

• Thresholds and budgets are in profile.yaml (fast_scoring:, llm:) and on the Settings page. Changing one never re-evaluates anything — jobhub rescore --recompute-only re-buckets instantly.
• Model-visible constraints (term, locations, work authorization, and profile.md) change the profile hash, so everything is re-scored. That's free. When a change only widens acceptance, jobhub rescore --carry-over keeps the deep evaluations you already paid for.
• Two work terms. term + term_also_accept is the primary track; alt_terms is a second one, scored the same way but shown under its own chip and digest section. Postings that state no term — most board postings — appear under both. Empty alt_terms to go back to one.
• Degree (degrees: in profile.yaml, or the Profile page): the degrees you hold or are pursuing. The highest is a ceiling — a posting is archived only when the lowest degree it accepts is above it, so "Intern, PhD, Quantitative SWE" and "MS/PhD" go while "BS/MS" and "BS, MS or PhD" stay. A PhD mentioned under "Preferred qualifications", in a pay band, or as "a plus" is not a requirement. Graduation dates and year of study are never a reason to reject. Free and not model-visible: changing it re-scores nothing, and jobhub rescore --recompute-only applies it to jobs already scored. Empty the list to switch it off.
• Editing prompts/ or schemas/ means bumping RUBRIC_VERSION in jobhub/config.py, then jobhub rescore.

Repo layout

jobhub/            pipeline: ingest, prefilter, fastscore, evaluate, digest, home
jobhub/web/        Flask UI (app.py, themes.py, static/, templates/)
profile/           your profile.md, profile.yaml, companies.yaml
prompts/ schemas/  what the model is asked and what it must return
data/ digests/     SQLite DB, logs and generated digests (gitignored)

Note: jobhub/web/templates/ is listed in .gitignore. The ten templates already committed stay tracked (gitignore doesn't apply retroactively), but new ones won't be picked up — and because pyproject.toml ships web/templates/*.html as package data, any template that isn't committed will be missing from a fresh clone or an install from a built distribution. Run git check-ignore -v <file> if a template seems to vanish.

Tests

.venv/bin/python -m pytest -q