Lohnt sich dein eigenes Gojiberry AI?

Build me a signal-triggered LinkedIn outreach agent to replace Gojiberry AI. Requirements:

- Local Node + TypeScript service: Express dashboard on localhost:3000, better-sqlite3
  for storage, node-cron for the loop. No frontend framework.
- I define my ICP once in icp.yaml: titles, company sizes, geos, and the competitor
  LinkedIn pages and creator profiles to watch.
- Every 6 hours, pull signals with apify-client (token in .env): likers and commenters
  on watched posts, new followers, and job changes. Upsert each person into a prospects
  table with the signal, its URL, and the date it fired.
- Score each prospect 0-100 against the ICP in one LLM call (key in .env) with a
  two-line reason. Under 70 is never contacted.
- Enrich everyone above 70 through one provider, MoltSets or Prospeo · pick whichever
  ships a Node client, and cache by profile URL so I never pay twice for the same person.
- Draft a connection note under 300 characters plus two follow-ups, written from the
  profile and the exact signal that fired.
- Send through unipile-node-sdk. Connect several LinkedIn accounts and round-robin
  across them at 20 invites and 40 messages per account per day, randomized gaps in
  business hours, invite first and follow-ups only after acceptance. Poll replies every
  15 minutes and stop the sequence the moment one lands.
- Drafts wait in an approval queue until I click Send · a --auto flag skips it. The
  dashboard lists prospect, signal, score, sender account, and thread. No accounts, no
  telemetry, everything on my machine except the Apify, enrichment, Unipile, and LLM
  calls.
- Out of scope: email sequences and a hosted control plane. Do not scrape LinkedIn
  directly, every LinkedIn action goes through Unipile.
- README: the Apify actors used, how to connect each LinkedIn account in Unipile, the
  .env keys, and a warning that per-account limits are real, so keep the caps low for
  the first two weeks.