Job Radar

Polls company job boards directly across 5 ATSs (Greenhouse, Lever, Ashby, Workday, SmartRecruiters), rules-filters then LLM-scores new postings against my profile, and pings a Discord channel with color-coded, urgency-tagged embeds. Runs free on GitHub Actions cron. Polling the ATS APIs directly means new reqs are seen within minutes of going live, upstream of LinkedIn/Indeed aggregators.

Currently watching 300+ live boards across big tech, fintech, and a deep bench of niche/YC startups (AI infra, dev tools, robotics). The live watch-list and scoring profile are private config (they encode my personal targeting); config/companies.example.yaml and config/profile.example.yaml show the shape, and CI fetches the real ones at runtime from a private repo. Grow your own list with seed + validate; dead slugs are skipped, and newly-added boards prime silently (no flood).

Quick start (local, no secrets needed)

python3.12 -m venv .venv && . .venv/bin/activate
pip install -e ".[dev]"
pytest -q                       # run the test suite
cp config/companies.example.yaml config/companies.yaml   # then make them yours
cp config/profile.example.yaml config/profile.yaml
python scripts/smoke.py         # hit one real board per ATS
python -m job_radar --preview --company openai   # see what would surface from one board
python -m job_radar --dry-run --limit 5          # full pipeline, console output, 5 boards

CLI (python -m job_radar)

Flag	What it does
--preview	Show what would surface from the current backlog, ranked by fit. Read-only (no priming, no state writes). Best for tuning profile.yaml.
--dry-run	Run the full pipeline but print to console instead of posting to Discord.
--prime	Mark everything seen without notifying (re-prime; e.g. after broadening the profile).
--backfill	One-time: score the current open backlog and write strong matches to the Sheet (no pings, no state change). Seeds the tracker with today's inventory. Reaches back --backfill-days (default 60, wider than the cron) and writes only --backfill-min-fit and up (default 60). Needs the Sheet env vars; works without an LLM key (heuristic).
--company SLUG	Only poll one company (local testing).
--limit N	Only poll the first N companies.
--profile / --companies / --state PATH	Override config/state paths.

Tools

python -m job_radar.seed <list.csv>      # bulk-add companies (dedups). See format below.
python -m job_radar.validate             # check every board live; report ok/empty/dead
python -m job_radar.validate --prune     # ...and drop dead slugs from companies.yaml

Seed CSV format (one per line; # comments):

slug,ats[,tier]                       # greenhouse | lever | ashby | smartrecruiters
slug,workday,tier,wd_host,wd_site     # workday needs the tenant host + site

Config

• config/profile.yaml — what counts as relevant to me (roles, locations, thresholds, and the free-text summary the LLM scores against). Tune with --preview.
• config/companies.yaml — the watch-list. tier: dream auto-pings (bypasses the LLM); reserve it for high-fit companies. tier: target lets the LLM filter. Banks are target so their broad co-op pools get narrowed to SWE/AI/Data.

Finding a Workday tenant: open the company's careers page and watch the network request to /wday/cxs/<tenant>/<site>/jobs. The host is the wdN part (e.g. cibc.wd3.myworkdayjobs.com → wd_host: wd3, wd_site: campus).

Going live (free, GitHub Actions)

1. Create a Discord channel + webhook; copy the URL.
2. Get an LLM key. Default scorer is Gemini Flash (free tier, gemini-2.5-flash). For Claude (Anthropic API), set LLM_PROVIDER=claude (defaults to Haiku 4.5). For AWS Bedrock (no per-day free-tier cap, best if you have AWS): pip install -e ".[bedrock]", set LLM_PROVIDER=bedrock, add AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_REGION secrets, and set LLM_MODEL to your enabled Bedrock model id (e.g. anthropic.claude-3-5-haiku-20241022-v1:0, or a us./eu. inference-profile variant for your region). Any scorer falls back to a free deterministic heuristic if it errors, so a quota blip never drops coverage.
3. gh repo create job-radar --private --source=. --remote=origin --push
4. Repo Settings → Secrets → Actions: DISCORD_WEBHOOK_URL, LLM_API_KEY, optional DISCORD_ROLE_ID, LLM_PROVIDER.
5. Trigger once via the Actions tab (confirm a PRIMED run — the first run primes silently so you aren't flooded), then the ~15-min cron takes over.

Tracker (Google Sheet + reminders, no server)

The cron doubles as an application tracker. When the Sheet secrets are present, each poll mirrors every match into a Google Sheet (one New row per job), and a second daily workflow (remind.yml) reads the Sheet and pings Discord with a catch-up: how many roles are pending and how long since you last applied, then must-apply (rows you flagged high priority), due soon (deadline within 3 days), and strong roles still unapplied. You triage in the Sheet — set Priority (must/high/med/ low), fill Deadline, mark Status Applied/Skip — and reminders stop nagging a row once it leaves New. So a busy stretch just means the backlog waits for you, sorted by what you said matters. No always-on process: it all rides the free Actions cron.

Two extra Action secrets turn it on (the poll skips the Sheet cleanly if they're absent):

• GOOGLE_SHEET_ID — from the Sheet URL, the part between /d/ and /edit.
• GOOGLE_CREDENTIALS — the full service-account JSON key (paste the file contents).

Google setup: enable the Sheets API, create a service account, download its JSON key, and share the Sheet with the service-account email (Editor). The Sheet's columns are created automatically on first write. Locally you can preview reminders with GOOGLE_SHEET_ID=... GOOGLE_CREDENTIALS_PATH=google-creds.json python -m job_radar.remind.

The Sheet otherwise fills only as new roles appear (everything already seen was primed). To seed it with today's open inventory immediately, run --backfill once (scores the current backlog and writes matches; no pings, no state change). Mind the LLM quota: the Gemini free tier is ~200 requests/day shared with the cron, so the backfill is capped (raise BACKFILL_CAP with a paid key).

Optional: always-on interactive bot

If you'd rather press Applied / Not for me buttons and use slash commands (/pending /top /due /stats) instead of editing the Sheet, run the bot as a persistent process (a VM or a machine that stays on). Same Sheet, plus Discord buttons. Design: docs/superpowers/specs/2026-06-07-tracker-bot-design.md.

pip install -e ".[bot]"
python -m job_radar.bot          # or: python -m job_radar.bot --check  (test the Sheet only)

Put the values in a git-ignored .env: DISCORD_BOT_TOKEN, DISCORD_CHANNEL_ID, GOOGLE_SHEET_ID, GOOGLE_CREDENTIALS_PATH=google-creds.json, LLM_API_KEY. Bot invite scopes: bot + applications.commands; permissions: Send Messages + Embed Links (Guild Install).

How it works

sources fetch all boards concurrently → dedup drops already-seen (and primes silently on first run, and per-board when you add a new company, so neither floods) → filters free rules pre-filter → scorer LLM-scores survivors (Gemini/Claude, enriching Workday/SmartRecruiter descriptions first; a deterministic heuristic transparently covers any posting the LLM can't score, e.g. on a rate limit, so an outage degrades coverage instead of dropping it) → urgency classifies (🔴 fresh+high or dream / 🟡 relevant / 🟢 weak→digest) → notify posts to Discord. State persists in the Actions cache; a weekly keepalive keeps the cron from auto-disabling. Design docs in docs/superpowers/.