Back to projects
AI AgentsWorkflow AutomationProduct ThinkingEvals (planned)

Job-search pipeline: an AI agent that finds, scores and alerts me to PM roles

Oct 2026

A scheduled AI agent that searches job boards four times a day, scores each role 0-100 against my profile, logs it to a tracker, and pings me on Telegram, with me deciding what to apply to. Built for one user: me.

Job-search pipeline: an AI agent that finds, scores and alerts me to PM roles

The problem

Searching for AI PM roles across LinkedIn, Indeed, company career pages and alerts meant opening a dozen tabs a day, losing track of what I'd seen, and missing fresh postings.

What I built

- A scheduled Claude agent runs four times a day (08:40, 12:40, 15:40 and 19:40 IST). It checks company career boards (Lever, Greenhouse, Ashby, SmartRecruiters), job-alert emails, Himalayas and Indeed.

- Each role gets a 0-100 fit score against my profile, a one-line "why it fits", a seniority tag and an estimated pay range, and is logged to a Notion tracker with a dedup key so nothing appears twice.

- Strong, fresh matches (score 70+, at most 8 per run) are sent to a private Telegram channel through Make.com.

- On demand, I message the bot "find me jobs" and it replies with the best open roles: not yet applied or skipped, score 60+, not senior-level, and paying roughly ₹12 LPA or more.

Product decisions

- A human decides. The system never applies. It surfaces and ranks; I choose and submit.

- Deterministic filters, model scoring. The model judges fit; plain rules (status, score, seniority, pay) decide what reaches me.

- The tracker is the source of truth. Marking a role Applied or Skipped removes it from future results.

- No scraping of sites that forbid it. LinkedIn, Naukri and Wellfound are covered through their alert emails.

What went wrong

- The first version of the bot showed senior roles and low-paying roles, so I added seniority and pay filters.

- The pay filter is a text pattern, so it can't see dollar or "not listed" pay, and those roles still show.

- Some "why it fits" lines are cut off where they contain an email address.

- If nothing matches, Make ends the run silently, so I added a "checking your tracker" message first.

What I haven't done yet

I haven't measured how good the scoring is. Next step: hand-label 30 roles as "would apply / wouldn't" and compare against the agent's scores. It's built for one user and has no outcome data such as interviews or offers.

Stack

Claude (scheduled agent), Notion (tracker), Make.com, Telegram bot, Gmail alerts.

Videos