Job Bidder：多租户 AI 求职代理流水线，抓岗、定制简历并追踪申请

Job Bidder

Multi-tenant AI job-application agency — scrape job boards, auto-tailor a resume per role, and track applications across admins, workers, and clients.

Admin clicks Refill Jobs → JobSpy scrapes live boards → jobs land in each client's queue → a worker opens a job and the AI tailors the client's resume to that exact JD → marks it applied → the client sees a read-only view with week-by-week history.

Built to run an agency at $0 — self-hosted Postgres, free-scrape boards, and an OpenAI-compatible model (e.g. freeinference.org, Ollama, LM Studio) — no paid CRM, no paid enrichment.

✨ Screenshots

Admin — Applications (every job across clients with filters, dedup, and statuses)

Admin — People & Clients (workers ↔ clients, assignment, quotas)

Admin — Resumes (per-client base resume upload, text extraction, and a one-click General preset generator that auto-derives ~10 fitting roles from each resume)

Admin — Settings (AI provider config, earnings rates, team, maintenance banner, scrape history, downloadable backups)

Worker — Queue (their client's jobs with Working/Applied/Skipped tabs, a weekly earnings pool meter, and a worker self-serve Refill button with loading + results feedback)

Worker — Completion History (applied jobs grouped by week)

Roles

Role	What they do
Admin	Manage clients + workers, run scrapes, upload base resumes, configure AI, reset passwords, broadcast maintenance notices, see everything
Worker	Tailor + apply their single assigned client's jobs, one at a time
Client	Read-only view of jobs applied on their behalf + week-by-week history

1 worker : 1 client, no rotation. Role checks are app-level (lib/auth.ts requireRole) — no DB RLS.

Features

Job scraping via JobSpy across 10+ boards (Indeed, LinkedIn, RemoteOK, WeWorkRemotely, Remotive, WorkingNomads, Greenhouse, SmartRecruiters, BuiltIn, ZipRecruiter) with per-board timeouts so one slow board can't stall a run.
AI resume tailoring — the LLM returns structured JSON; rendered into a clean A4 PDF per job via react-pdf (@react-pdf/renderer). Downloadable as a proper .pdf.
Policy-aware search — search by title, company, board, or location; status + client filters.
Role-specific UIs — dense tables on desktop, compact cards on mobile.
"Continue where I left off" — DB-backed last-viewed cursor so a worker resumes exactly where they stopped (survives refresh / devices).
Weekly completion history for workers and clients.
Maintenance / announcement banner the admin can toggle for downtime or notices.
Encrypted AI keys (AES-256-GCM), local-disk resume/PDF storage, auth-gated file serving.

Quick start (local dev)

Prereqs: Node 18+, Docker (local Postgres) or a reachable Postgres, Python 3 for real scraping.

npm install
pip install jobspy pandas          # optional: enables real scraping

# 1. Postgres
docker compose up -d db             # host port 5434
#    or set DATABASE_URL to your own Postgres

# 2. Apply schema + seed accounts
npm run db:migrate
npm run db:seed

# 3. Run
npm run dev

Open http://localhost:3000 → /login. Default admin: admin@jobbidder.com / changeme (set ADMIN_PASSWORD before seeding for real use). Use Settings → AI Configuration to add a model key.

Environment (`.env.local`)

DATABASE_URL=postgres://jobbids:***@localhost:5434/job_bidder
AUTH_SECRET=...                    # npx auth secret
AUTH_TRUST_HOST=true
ADMIN_EMAIL=admin@jobbidder.com
ADMIN_PASSWORD=change-me
STORAGE_DIR=./data/uploads          # gitignored

# AI (fallback only — configure in-app at Settings → AI)
AI_STUB=true                       # true = deterministic demo output, no key
# AI_PROVIDER=custom                # openrouter | anthropic | deepseek | custom
# AI_MODEL=deepseek-v4-flash
# AI_BASE_URL=https://freeinference.org/v1
# AI_API_KEY=...

Other providers: openrouter (OPENROUTER_API_KEY), anthropic (ANTHROPIC_API_KEY), deepseek/custom (base URL + key). All keys are encrypted at rest in app_config once set in-app.

Architecture

app/
  (auth)/login/          real email+password (Auth.js credentials+JWT)
  api/auth/[...nextauth] Auth.js handler
  admin/dashboard/       4 tabs + Refill Jobs (scrape) + hooks/useJobs
  worker/queue/          worker client's job queue (+ continue-where-I-left-off)
  worker/job/[id]/       2-column detail: job info + Tailor | Answer | Submission
  worker/history/        weekly completion history
  client/jobs/           read-only applied jobs
  client/history/        client app history grouped by week
  api/
    scrape/              spawns scripts/run_jobspy.py, dedup, insert
    tailor/              AI -> structured JSON -> react-pdf -> stored .pdf
    files/[...path]      auth-gated resume / tailored-PDF / proof serving
    jobs, profiles, users, config, snippets, upload, upload-proof, pdf
db/    pool.ts  schema.sql  migrations/ migrate.ts seed.ts repo.ts
lib/   auth.ts  ai.ts  resume-pdf.tsx  storage.ts  crypto.ts  types.ts
components/   DashboardLayout (sidebar), JobTable, Modal, StatusBadge, panels
scripts/run_jobspy.py  JobSpy runner (per-board timeout, JSON-safe output)

JobSpy + scraping

Refill Jobs spawns scripts/run_jobspy.py <config> <out.json>, which calls the jobspy library per (term, site) with a per-board deadline (so LinkedIn blocking can't hang a run), sanitizes NaN/Infinity, and Node dedupes by URL per profile before inserting.

Deploy (VPS / PM2 + Cloudflare tunnel)

apt install postgresql
sudo -u postgres createuser -P jobbids
sudo -u postgres createdb -O jobbids job_bidder

npm install && pip install jobspy pandas
npx auth secret
# set DATABASE_URL, ADMIN_PASSWORD, and in-app AI config
npm run db:migrate && npm run db:seed && npm run build

pm2 start npm --name job-bidder -- start
pm2 start cloudflared --name job-bidder-tunnel -- tunnel --url http://localhost:3000
pm2 logs job-bidder-tunnel        # the public trycloudflare URL

Backups

backup.sh dumps the Postgres DB + uploads into one dated .tar.gz; restore.sh brings it back. Runs daily via cron. Copy archives off-server.

What it deliberately doesn't do

No auto-email-reply sourcing, multilingual, multi-language resumes/version history, chat, calendar, or notifications. Status transitions are manual (worker-driven).