Skip to content
Shivam Goel
← All the work

Automation · Personal, 2025

A GTM pipeline I built for myself

A daily job that scans public ATS boards, scores every role against my resume with an LLM, drafts a tailored cover note for strong fits, and emails me a digest. Same shape as any outbound engine, except the ICP happens to be me.

What came out of it

5
Job sources unified into one queue
Daily
Runs unattended on GitHub Actions
0
Auto-submitted applications, review is mandatory

Context

Job hunting is an outbound problem in reverse: a noisy top of funnel, a scoring step, a personalisation step, and a review queue. I would rather build the machine once than do the work daily.

SOURCES Greenhouse Lever Ashby Workable RemoteOK normalise + dedupe SQLite filter remote/YoE deterministic score vs. resume LLM draft note strong fits LLM email digest human review no auto-submit DAILY · GITHUB ACTIONS · CHEAP FILTERS BEFORE EXPENSIVE JUDGEMENT
DIAGRAMfetchers → normalise/dedupe → filter → score → draft → digest

What I built

  1. Board API fetchers

    Greenhouse · Lever · Ashby · Workable · RemoteOK. No scraping, no LinkedIn.

  2. Normalise and dedupe

    A local database, so a role is never surfaced twice

  3. A deterministic prefilter

    Remote, seniority, title. Runs before any model call, so the expensive step only sees candidates worth scoring.

  4. LLM scoring

    Against a rubric built from my actual resume, with a tunable fit threshold

  5. Drafting for strong fits only

    A tailored cover note, then a single digest email

  6. Scheduled on GitHub Actions

    Draft mode by design. Nothing is ever submitted without me reading it.

What I'd do differently

The deterministic filter before the model call was the decision that made it viable, because scoring every role with an LLM would have cost more than it saved. Cheap filters first, expensive judgement last, is a rule I now apply to most pipelines.