Daily Job Matcher

Daily Job Matcher

An end-to-end job-search assistant: scrape → knockout-gate → score → curate → Notion, plus email → status updates. The mechanical steps run as a tested Python package (jobscreener); the judgement steps (reading full job descriptions, curating true fit, writing to Notion) are done by you with the Notion/Gmail MCP connectors.

When to use

Trigger for: "find me jobs that fit my resume", "screen these LinkedIn roles", "today's job matches", "which of these should I apply to", "update my application tracker from my email", "did I get any rejections", "build a daily job digest". Prefer this over ad-hoc searching whenever a résumé + real postings are involved.

One-time setup

1. Install: pip install -r requirements.txt (add python-jobspy only if

using the free JobSpy source; on Python 3.14 use pip install --only-binary=:all:).

1. Secrets in .env (copy .env.example; git-ignored — never commit or paste):

APIFY_TOKEN (required for the LinkedIn source), optional ANTHROPIC_API_KEY (LLM re-scoring), NOTION_TOKEN for the unattended path.

1. Résumé: put the user's CV at data/cv.md (copy data/cv.example.md). This

is the matching source of truth — level, skills, target roles, and hard constraints (visa, years, degree).

1. Search config: apify_input.json (copy apify_input.example.json) — the

Apify actor input: keyword[], locations[], publishedAt (r86400 = 24h), maxItems. Set JOBSCREENER_APIFY_ACTOR / JOBSCREENER_APIFY_TASK in .env.

1. Candidate constraints: tune CandidateProfile in src/jobscreener/config.py

(years_experience, max_years_experience, needs_visa_sponsorship, highest_degree, target_roles, exclude_title_keywords, …).

1. Notion: connect the Notion MCP and note the target database's

data_source_id. Use one master table (see step 4 of the daily cycle).

Daily cycle

Run from the project directory. (If python is a broken Windows Store stub, use py or the full interpreter path.)

1. Scrape the recent window (Apify actor; ~$0.1/run):

python -m jobscreener run --source apify --apify-input apify_input.json → writes output/apify_raw.json.

1. Screen + dump full JDs. Apply the pipeline (knockout gate + score) to the

raw file and export the passing candidates with full descriptions so you can read them. The heuristic rank is a first pass, not the final word.

1. Read the full JDs and curate by true fit. The knockout gate already drops

ineligible / over-experienced / wrong-degree roles; your job is judgement: down-rank roles that only keyword-match (e.g. trading/HFT, AI-research engineering, pure SWE) but don't fit the candidate, and flag caveats the gate can't see. Quality over quantity — a short honest list beats a padded one.

1. Append to the ONE master Notion table (via the Notion MCP). Do not

create a new table each day. Columns: Name, Company, Location, Fit (Strong/Good/Moderate), Notes (why / caveats), Seniority, Status (new matches = To Apply), URL, Date, Source (LinkedIn). Prefix Name with a priority number. UPSERT rule — never insert before searching. Query the tracker for BOTH a distinctive company token AND a title fragment; use the shortest distinctive token so name variants still match ("Agricole" not "Crédit Agricole CIB HK" — accents, suffixes, and HK/SG tags have all defeated exact matching before). If any plausible match exists, UPDATE that row (or skip) — do not add a second.

1. Report the ranked shortlist and which to apply to first.

Use --exclude-seen so the same job isn't surfaced on consecutive days.

Knockout gate (the core idea)

Standard job-matching separates hard non-negotiables (disqualify instantly) from soft weighted signals (rank). Implemented in filters.py:

• Eligibility — citizenship-only / no-sponsorship / clearance phrases → out.
• Experience — the JD's minimum required years > max_years_experience → out.

("Minimum 3 years" means ≥3; a new grad doesn't meet it.)

• Degree — PhD / postdoc required and the candidate is below → out.

Soft score (0–100) blends skill match (TF-IDF + skill overlap), title relevance, domain fit, seniority alignment, and location. Full rationale: METHODOLOGY.md.

Application tracking from email

When asked to "read email / update statuses": search Gmail (Gmail MCP) for job-application confirmations, rejections, and interview invites, then update the matching rows' Status in the same master table (Applied / Rejected / Interview / Offer / Started), or add rows for applications not yet tracked (Source = Email). One table for both matching and tracking. Backfills follow the same UPSERT rule as step 4: before adding a row from an email, search the tracker by a short company token + title fragment — the daily cycle may already have added the same role under a slightly different company name (this produced real duplicates on 2026-07-08: CACIB, Milliman). Match found → update its Status/Notes; only add when nothing plausibly matches.

References (read as needed)

• RUNBOOK.md — full operating manual, gotchas, module map, how to extend.
• METHODOLOGY.md — how the funnel and scoring decide (the research).
• README.md — install, sources (Apify / JobSpy / sample), CLI flags.
• src/jobscreener/ — the package; tests: python -m unittest discover -s tests -t ..

Notes

• No in-house LinkedIn scraper (ToS). Sources are pluggable behind JobSource

in sources.py: Apify (managed, recommended), JobSpy (free), sample/CSV/JSON.

• The Notion/Gmail MCP connectors are verified to work in headless scheduled

runs (2026-07-03); the cycle runs as the nightly daily-job-feedback-sync scheduled task AND on command. If a headless run can't reach a connector, report-and-stop rather than fail silently.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: chenxi-bot21
• Source: chenxi-bot21/ai-job-search-pipeline
• License: MIT

Install and usage instructions live in the source repository linked above.