linkme/README.md

linkme

A personal job-hunting pipeline for the Swiss market. It scrapes LinkedIn (plus a handful of Swiss boards), filters out everything that doesn't fit the profile in cv/polo.md, appends what survives to a CSV, and pushes the good ones to Telegram twice a day.

Two things run on their own once installed:

automation	when	what it does
daily digest	08:00 + 18:00	fetch every search, push new profile-matching jobs to Telegram
sciCORE watch	08:47	alert if sciCORE's open-positions page lists a software role

Quick start

bun fetch      # scrape all searches, append new jobs to results/jobs.csv
bun daily      # fetch, then Telegram the new jobs that match the profile
bun recheck    # show what the filters are currently dropping + keyword suggestions

bun fetch needs no setup — reruns are safe, already-seen jobs are skipped by ID. Telegram needs the two-line setup below.

Telegram setup

# 1. Message @BotFather → /newbot → copy the token into .env
echo 'TELEGRAM_BOT=<token>' >> .env

# 2. Send your bot any message ("hi"), then discover your chat ID
bun run telegram.ts --setup      # prints TELEGRAM_CHAT_ID=... — paste it into .env

Then bun daily --dry-run prints the digest instead of sending it, and bun daily --resend re-sends today's jobs from the CSV without re-fetching.

.env also holds LI_AT, a LinkedIn session cookie used only for fetching individual profile pages. Never send it with a job search — with the cookie LinkedIn serves a React SPA instead of server-rendered HTML and the card parser gets nothing. .env is gitignored.

Structure

config.ts          # keyword families, search areas, the sweep() generator
fetch.ts           # LinkedIn scraper, HTML parser, all title/content filters
fetch-other.ts     # extra sources: itjobs.ch, job-room.ch, stellen-basel.ch,
                   #   datacareer.ch (403, currently dead), Basel intl orgs
index.ts           # orchestrator: run searches → dedupe → content-check → CSV
recheck.ts         # raw-mode fetch: lists what the filters dropped
daily.ts           # digest: fetch, match against the profile, send to Telegram
profile.ts         # title-level scoring — decides what's worth a notification
telegram.ts        # Telegram delivery + --setup chat-ID discovery
scripts/           # one-off searches (*-temp.ts) and scicore-watch.sh
cv/polo.md         # the profile everything is ranked against
results/jobs.csv   # output, one row per job
results/sent.txt   # links already pushed — never sent twice

How a job gets through

Two stages. Cheap title filters first, then one page fetch per survivor.

Title filters (inside fetchJobs, no network cost):

1. Switzerland guard — the card's location must say Switzerland. LinkedIn ignores geoId on unauthenticated requests and will happily return global results.
2. NON_ENGLISH_RE — German gender markers (m/w/d, /in), umlauts, French H/F and accents.
3. JUNIOR_TITLE_RE — junior, intern, trainee, werkstudent, entry-level.
4. JAVA_TITLE_RE — Java-titled roles.
5. BIO_DOMAIN_RE — bioinformatics, biostatistics, genomics, GMP/MSAT, bench-scientist titles. Filters on the role, not the employer: software and data jobs at Roche, Novartis or Swiss TPH stay in.
6. SOFTWARE_TITLE_RE — requires a real tech/role term. Skipped in raw mode, which is what bun recheck uses.
7. GERMAN_TITLE_RE — "(German-speaking)", "deutschsprachig".

Content check (index.ts only, one fetch + 350 ms per job): isViableContent counts German/French indicator words and drops at ≥ 6, then requiresGerman() drops English postings demanding German above B1.

That last one is deliberately more than a keyword match. For each mention of German it takes the nearest qualifying marker, and ignores a marker when another language name sits between it and the mention — both rules came from real false results, like "Fluent in English; German and/or French advantageous", where "Fluent" sits closer to "German" than "advantageous" does but qualifies English.

Profile match (profile.ts, digest only) scores the title against the CV's stack and seniority signals, needing 4+ points to notify. Hard rejects encode standing decisions: Microsoft stack, Java/C++/SAP, life-science domains, non-software functions, and the blocked companies Jobgether (an aggregator, not an employer) and ALLPS (postings require German despite passing the English filter).

Jobs are deduped globally by ID against results/jobs.csv on startup, so nothing is ever fetched or sent twice.

Output

results/jobs.csv — date_found, posted_at, search, title, company, link, source. Open it in any spreadsheet and delete rows as you apply or dismiss.

Configuring searches

config.ts holds four keyword families as separate OR-strings — KEYWORDS, PLATFORM_KEYWORDS, RESEARCH_KEYWORDS, FIELD_KEYWORDS. They're kept apart on purpose: LinkedIn ranks a focused query far better than one 50-term disjunction. sweep() generates one search per area per family (Research skips Bern; Platform and Field add a Switzerland-wide remote pass).

export const searches: SearchConfig[] = [
  {
    label: "Basel (city)",
    keywords: KEYWORDS,
    geoId: "103383283",
    workType: ["onsite", "hybrid"],
    timePosted: "month",
  },
];

SearchConfig also takes jobType, seniority, and location (use "City, Switzerland" when the geoId is unknown — LinkedIn resolves by name). A url: field overrides all filter params, so you can paste a LinkedIn search URL directly.

Known Swiss geoIds:

area	geoId
Basel (city)	103383283
Basel Metropolitan Area	90009885
Basel-Country	107117530
Zürich (city)	102436504
Zürich Metropolitan Area	107814425
Zürich Area	90009888

One-off searches

Don't edit config.ts for a single search — write a throwaway in scripts/:

import { KEYWORDS } from "../config.ts";
import { fetchJobs } from "../fetch.ts";

const jobs = await fetchJobs({
  label: "Luzern", keywords: KEYWORDS,
  location: "Luzern, Switzerland",
  workType: ["onsite", "hybrid"], timePosted: "month",
});
for (const j of jobs) console.log(j.title, j.company, j.link);

bun run scripts/luzern-temp.ts

*-temp.ts scripts are disposable — delete them when the search is done.

Scheduling

Both automations are launchd agents, not cron — this machine has no cron at all.

launchctl list | grep -E 'linkme|scicore'                     # are they loaded?
launchctl unload ~/Library/LaunchAgents/com.rafapolo.linkme-daily.plist   # stop the digest
launchctl unload ~/Library/LaunchAgents/ch.polo.scicore-watch.plist       # stop the watcher

If the Mac is asleep at the scheduled time, launchd runs the job once it wakes rather than skipping the day.

Logs: results/daily.log for the digest, results/.scicore-watch/watch.log for the watcher. The watcher logs every run — no change, change ignored, or the diff it sent — so silence on your phone can always be told apart from a broken script. Test its alert path without waiting for the real page to change:

SCICORE_URL="file://$PWD/fixture.html" bash scripts/scicore-watch.sh

recheck

bun recheck re-runs every search with SOFTWARE_TITLE_RE bypassed and lists each job the filters would drop, plus a word-frequency table of those titles. Use it to spot false negatives; ignore the generic suggestions (engineer, development, senior) and look for real role terms worth adding to the regex.