Internship Tracker

Polls Greenhouse and Lever job boards for tracked companies, stores postings in Supabase, and ranks internships against target roles using sentence embeddings.

Live: frontend · API

Currently tracking 63 companies and ~10,200 open postings, of which ~120 are internships.

Architecture

Three independently deployed pieces:

Vercel          React + TypeScript SPA          frontend/
  │ VITE_API_URL
  ▼
Render          Flask API + MiniLM scoring       backend/api.py, backend/scoring.py
  │
  ▼
Supabase        Postgres                         db/schema.sql
  ▲
  │
GitHub Actions  poller, every 30 min             backend/poller.py

The poller and the API are separate entry points against the same database. The poller deliberately does not import backend.scoring — see Why two requirements files.

Local development

Backend:

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python -m backend.api    # http://127.0.0.1:5000

Frontend:

cd frontend
npm install
npm run dev              # http://localhost:5173

.env in the repo root, for the backend:

SUPABASE_URL=...
SUPABASE_KEY=...

The frontend needs no config locally — it falls back to http://127.0.0.1:5000.

Run the poller by hand:

python -m backend.poller

API

Method	Route	Returns
GET	/companies	every company with status='active'
POST	/companies	verifies a name against Greenhouse then Lever, inserts on a hit. {message, source}, or 404 if on neither
GET	/internships?limit=50	internship-titled open postings, newest first
GET	/postings	all open postings — ~10k rows, ~5MB. Unused by the frontend
POST	/targets	{targets: string[]} → postings scored by cosine similarity, descending

Layout

backend/
  api.py                 Flask routes
  poller.py              board polling, upserts, closure detection
  db.py                  every Supabase query
  scoring.py             MiniLM embeddings + cosine similarity
  titles.py              is a title an internship? (no heavy imports)
  boards.py              Greenhouse/Lever API calls
scripts/
  add_companies.py       one-off seeding
  prefetch_model.py      build-time model download
db/
  schema.sql             current schema
  migrations/            ordered migrations

frontend/src/
  App.tsx                tab state, composes features
  features/              FindRoles, AddCompanies
  hooks/                 useInternshipSearch, useCompanies — state lives here
  components/            presentational only
  api/client.ts          transport: base URL, ApiError, request<T>
  api/index.ts           the four endpoint functions
  types.ts               API shapes

npm run typecheck && npm run lint && npm run build before pushing — vite build does not typecheck, so a type error will build and deploy fine.

Deployment

Vercel (frontend). Root Directory must be frontend; leave every command override off. Set VITE_API_URL to the Render URL. Vite inlines it at build time, so changing it requires a redeploy, not a restart.

Render (API). 512MB free tier is enough — measured ~200MB peak.

• Build: pip install -r requirements.txt && python -m scripts.prefetch_model
• Start: gunicorn backend.api:app --bind 0.0.0.0:$PORT --workers 1 --timeout 120
• Env: SUPABASE_URL, SUPABASE_KEY

--workers 1 matters: each worker loads its own copy of the model. --timeout 120 matters: /targets can exceed gunicorn's 30s default.

GitHub Actions (poller). .github/workflows/poll.yml, needs the same two Supabase secrets. Free on public repos. Note GitHub disables scheduled workflows after 60 days of repo inactivity.