LinkedIn Opportunity Radar
Finds posts where someone describes a need I can solve, scores them, drafts the reply and learns from what I do next.
- 0
- API keys and cloud calls
- 0
- messages it sends by itself
- 6
- labels a lead can carry
- 5+
- channels: LinkedIn, Reddit, Hacker News, feeds, a bookmarklet
01Summary
The Radar reads posts and looks for the ones that reveal a real need. It scores each, drafts a pitch and puts it in an inbox for me to triage. The analysis runs on the model built into the Mac. Nothing leaves the machine.
It never posts, comments, connects or messages. Every action in the outside world is mine.
02Why I built it
At work I had said plainly that I did not want to do outreach. What I actually disliked was the search: hours of scrolling to find three people worth writing to. So I built the part I disliked and kept the part that needs a person.
03How it works
- Step 1
Scan
Run only when I ask. A capped number of posts per scan.
- Step 2
Triage
The on-device model decides whether a post shows a need.
- Step 3
Score
A rule score, blended with a learned score as labels accumulate.
- Step 4
Draft
A pitch grounded only in past projects I have marked as safe to reference.
- Step 5
Label
Eligible, contacted, replied, converted, or lost and dismissed with a reason.
- Step 6
Learn
Each label retrains a small model. A backtest rejects a retrain that makes it worse.
04Decisions
| Decision | Reason |
|---|---|
| Nothing runs unattended | A scheduler exists and is not installed. I press the button. |
| Anonymous clients are never named | The drafter can cite only projects marked as referenceable. Everything else is invisible to it. |
| Replies, not pitches, on forums | Leads from Reddit and Hacker News get a helpful public reply suggestion. Anything about my own product must carry a disclosure. |
| Rollback and reset | If the learned model drifts, I can return to an earlier one or start again. |
05What broke
- Automating LinkedIn is against its terms and can restrict an account. The documentation says this first, scans are capped, and it stops at any verification prompt.
- The on-device model is small and misses things. The learning loop and my edits are what make it useful.
- Page selectors are best effort and need adjusting after real runs.
06What is next
More labelled leads, so the learned score can carry more weight than the rules.