auto-apply

A local job-search pipeline for new-grad and internship SWE roles. It watches public job lists and ATS boards, surfaces new postings the moment they appear, writes tailored application drafts keyed to each job's real form questions, and can auto-fill and submit applications through Playwright.

Everything runs on your machine. The database, profile, resume, and application history stay local.

What it does

Stage	What happens
Ingest	Pulls 8 GitHub-maintained job lists (SimplifyJobs, speedyapply, NUFT quant repos, etc.), dedupes across sources, filters to SF / NYC / Remote-US SWE and quant roles
Scout	Polls company ATS boards directly (Greenhouse, Lever, Ashby, SmartRecruiters, Workable), discovers YC startup boards, scans HN "Who's Hiring", and checks Remotive for remote early-career roles
Blocklist	Drops mega-cap / big-tech companies at ingest time via config/company-blocklist.json (Amazon-tier and above)
Enrich	Follows redirect URLs, detects the real ATS, fetches the job description, and pulls the actual application-form questions
Draft	Claude writes a cover letter plus an answer for every form field, using your profile and canonical screening answers
Apply	Playwright opens a headed browser, reads the live form, fills every field, attaches your resume, and submits (or stops before submit if AUTO_SUBMIT=0)
Notify	macOS notification when new jobs land: auto-apply: 3 new jobs — Nuro — SWE New Grad, +2 more — 3 drafted

The inbox is sorted by release time (firstSeenAt), not when you happened to open the app. A 870-job backfill on first run is normal; focus on New today and star/skip from there.

Stack

Next.js 15 (App Router) · React 19 · Tailwind v4 · Drizzle + SQLite · Anthropic SDK (structured outputs) · Playwright · launchd watcher

First-time setup

pnpm install
pnpm db:push                                # creates data/db.sqlite from schema
cp .env.local.example .env.local            # add ANTHROPIC_API_KEY
cp data/profile.md.template data/profile.md # fill in your background
cp resume/resume.pdf data/resume.pdf        # or use your own PDF
pnpm ingest                                 # first run = silent backfill of current listings
pnpm watcher:install                        # optional: launchd ingest every 15 min
pnpm dev                                    # http://localhost:3000

The UI shows a setup banner until .env.local, profile.md, and resume.pdf exist.

One-time config files:

• data/profile.md — your background; drafts are grounded in this
• data/contact.json — name, email, phone, LinkedIn, etc. (used by auto-apply for identity fields)
• config/screening-answers.json — canonical answers for work auth, comp, start date, EEOC prefs (used verbatim on select fields)
• config/company-blocklist.json — companies to exclude entirely (normalized name matching)

For a training-wheels period, set AUTO_SUBMIT=0 in .env.local. The browser fills everything but leaves the final Submit click to you.

How the pipeline works

Every 15 minutes (launchd watcher) or on Refresh in the UI:

1. Ingest fetches all source READMEs. Rows are filtered by title (SWE/quant, no sales/HR), location (SF, NYC, remote US, or quant hubs like Chicago/Austin for quant firms), and company blocklist. Jobs explicitly marked closed (🔒) in a source are closed immediately; jobs missing from every source for 24h+ are auto-closed.

2. Scout runs four keyless discovery streams:

   • ATS boards: polls boards for every known company (~140+ boards, thousands of postings per run). Catches postings before GitHub lists update.
   • YC discovery: probes actively-hiring YC companies for Greenhouse/Lever/Ashby boards (20 companies per run, cached in data/scout-state.json).
   • HN: "Who is hiring" comments plus front-page YC job posts (Algolia + Claude extraction, at most 2×/day).
   • Remotive: remote-US software-dev feed, early-career filtered (at most 2×/day).

   Two admission tiers: titles that say new-grad/intern/junior get notify + auto-draft. Generic SWE titles at startups enter quietly ("ambient": visible in inbox, no API spend). A board's first poll seeds quietly; only deltas from already-watched boards count as new releases.

3. Delta detection: a job is "released" the first time it appears in any source. The first-ever run (or any insert of 50+ jobs) is treated as backfill: stored, but no notifications or drafts.

4. Enrich + draft (new releases only): follow redirects, detect ATS, fetch JD, fetch real form questions (Greenhouse API; canonical fallback for others), then Claude drafts cover letter + per-field answers.

5. Notify via macOS osascript.

Auto-apply flow

When you click Apply on a drafted job (or run pnpm apply:batch):

1. Ensures a draft exists (enriches + tailors if missing)
2. Double-apply guard: warns if you already applied to the same company
3. Launches Playwright (headed by default)
4. Reads whatever form the page actually renders (live DOM extraction)
5. Fills fields from contact.json → tailored draft → Claude for anything unmatched
6. Attaches data/resume.pdf
7. Submits (unless AUTO_SUBMIT=0) or leaves the window open on CAPTCHA / unexpected state

Workday and LinkedIn are login-walled and stay manual. Every attempt is recorded in the applications table with per-field audit trail.

Commands

pnpm dev                 # UI at http://localhost:3000
pnpm ingest              # run the full pipeline once (what the watcher runs)
pnpm watcher:install     # launchd agent, every 15 min (logs: data/watch.log)
pnpm watcher:status
pnpm watcher:uninstall
pnpm blocklist:sweep     # retroactively skip open jobs from blocklisted companies
pnpm apply:dry <jobId>   # fill a form without submitting (sanity check)
pnpm apply:batch         # batch apply to drafted jobs (--limit, --gap, --starred)
pnpm db:studio           # browse the DB

Resume (LaTeX)

Source lives in resume/resume.tex (Jake's Resume template). Compile with:

cd resume && tectonic resume.tex

Copy the output to the apply path when ready:

cp resume/resume.pdf data/resume.pdf

Sources

Configured in lib/ingest/sources.ts:

• SimplifyJobs/New-Grad-Positions
• SimplifyJobs/Summer2026-Internships
• speedyapply/2026-SWE-College-Jobs
• ambicuity/New-Grad-Jobs
• zapplyjobs/New-Grad-Jobs-2026
• jobright-ai/2026-Software-Engineer-New-Grad
• northwesternfintech/2026QuantInternships
• northwesternfintech/2027QuantInternships

Project layout

app/                    Next.js UI (inbox, job detail, API routes)
db/                     Drizzle schema
lib/
  pipeline.ts           ingest → scout → enrich → draft → notify
  ingest/               README fetchers, parsers, dedupe, blocklist, NUFT quant parser
  scout/                ATS board polling, YC discovery, HN, Remotive
  enrich/               per-job: finalUrl + ATS + JD + form questions
  tailor/               Claude draft generation (structured outputs)
  apply/                Playwright auto-apply (live form extraction + fill)
  ats/                  ATS detection + form-question fetching
  jd/                   JD fetchers
  notify.ts             macOS notifications
config/
  screening-answers.json
  company-blocklist.json
data/                   local only (gitignored): db, profile, resume, contact, logs
resume/                 LaTeX resume source + compiled PDF
scripts/
  run-ingest.ts         pipeline CLI
  watcher.ts            launchd install/uninstall
  batch-apply.ts        batch auto-apply
  dry-run-apply.ts      fill without submit
  sweep-blocklist.ts    retroactive blocklist sweep

Notes

• Tailoring defaults to claude-opus-4-8 with prompt caching. Override with ANTHROPIC_MODEL in .env.local.
• Drafts and apply answers never use em dashes (style rule in the system prompt).
• Outbound fetches timeout at 20s. Enrichment runs in parallel batches of 5 so one slow career site cannot stall the pipeline.
• The watcher only runs ingest/scout/enrich/draft/notify. It does not auto-apply in the background. Applying always requires you to trigger it.

Roadmap

• now: watch → notify → real-form tailored drafts → Playwright auto-apply (Greenhouse, Lever, Ashby, most custom ATSes)
• next: per-job tailored resume variants from the LaTeX source
• later: contact discovery, follow-up tracker