Job hunting is itself a repetitive pipeline. The same experience, re-explained for every company; the same question, re-answered for every form. I wanted to measure one thing: hand the whole pipeline to a machine and find out how hard a cold-start job search actually is. Job Radar runs the entire process for me, and then I watch what the real world sends back. Not to spray applications — to measure.

The Methods

1. One source of truth about myself, maintained once. Experience, projects, boundaries, reusable answers — all in one candidate_kb. Every application’s material, form answers, and pitch derive from it. I stop deciding per application.
2. Automated discovery and scoring. Scan AI labs, startups, fellowships, and builder roles across Greenhouse and Ashby boards, plus founder posts on HN and communities; score against my real positioning; stage the high-fit ones into a queue.
3. Gated automated outreach. A Gmail pipeline capped at 30 a day, claim-send-record in three steps with live bounce and reply monitoring; founder DMs and ATS applications run in parallel.

The Result (Real Numbers)

1. The scale was real. The system reviewed and scored 29,007 leads across 24,553 companies and mined 36,256 contact channels. On the discovery end, the machine works.
2. The volume was real too — but the bounce rate was high. It sent 324 cold emails to 320 people. 93 of them hard-bounced — roughly 29% never landed.
3. Almost nothing came back. Of the ~230 that did land, replies from actual humans were in the single digits: one positive, one to follow up, one rejection. Call it a 1% human-reply rate, with positive replies near zero. The safe sendable pool then drained to zero effective unsent, and the volume stopped.

What’s Worth Watching Next

1. From cold volume to high-intent channels. Stop refilling a pool just to hit 30 a day; go where the project and the founder’s intent are visible — YC/Wellfound/small-team roles, founder and project-owner channels, Bonjour-style direct applications.
2. Tailored proof-of-work outreach. One precise direct application beats dozens of generic sends; only target where intent and my positioning — hands-free agent, voice, ADHD workflow — clearly line up.
3. The machine moves to the background; the human only judges. Automation stays on for monitoring; the system handles the mechanical part, and the human only decides whether a target is worth tailoring for.

The takeaway: a machine can push out volume, but it can’t manufacture someone else’s demand. The bottleneck was never scale. It was hit-rate and channel.