Hiring Signals — examples

Working code for Hiring Signals, an Actor that reads a company's public job board and returns what the board reveals instead of the listings on it.

Everything here is standard library only. Nothing to install.

---

What the tool actually returns

Point it at a company. It reads the board and answers one question: what is this company about to do?

Reddit · 188 open roles
  ↳ Exactly one legal role on the entire board, and it is a
    Director of Ads Policy.
    You do not hire a policy director for an ads business you
    are not about to scale.

Duolingo · 64 open roles
  ↳ 5 leadership openings in marketing, plus single roles in
    Japan, the UK and China while 78% of the board sits in the US.
    Three markets opening at once.

Notion · 110 open roles
  ↳ Half the board is go-to-market.
    The product is finished. They are buying distribution now.

Those are real, from one run. The scripts in examples/ are how you get them, and FINDINGS.md has a full pass over 112 company boards and 11,969 open roles — including the loudest signal of that run, which turned out to be wrong and why.

---

Examples

run_and_export.py	Run over a list of companies, write one CSV row per signal. The shape you want for a spreadsheet or a CRM import.
watch_target_accounts.py	Run on a schedule and print only what changed. The first run stores a baseline; after that you hear about a company the week it crosses a line it had not crossed before.

export APIFY_TOKEN=your_token

printf 'stripe\nfigma\nwise\n' > companies.txt
python examples/run_and_export.py companies.txt signals.csv

---

Input

{
  "companies": [
    { "board": "stripe" },
    { "board": "wise", "name": "Wise" },
    { "board": "https://nvidia.wd5.myworkdayjobs.com/NVIDIAExternalCareerSite", "name": "NVIDIA" }
  ],
  "minSeverity": 50,
  "includeJobs": false
}

board is the company's slug on its ATS — usually just the company name. The platform is detected automatically and then remembered, so a repeat run costs one request per company.

Seven platforms: Greenhouse · Lever · Ashby · Workable · SmartRecruiters · Recruitee · Workday

Workday is the exception to the slug rule: every customer gets its own host, so paste the careers page URL and it is recognised on sight.

---

The eight signals

GTM_PUSH	Revenue hiring outweighs product hiring. The product is ready; they are buying distribution.
GEO_EXPANSION	First roles in a country they have none in. The earliest visible sign of market entry.
FIRST_HIRE_IN_FUNCTION	A function being created from nothing.
AI_CAPABILITY_BUILD	Concentrated ML hiring. They are building AI, not buying it.
INFRA_SHIFT	First platform and SRE hires. The original stack stopped working.
LEADERSHIP_BUILD	Several director-and-above searches in one function. An org is being built.
HIRING_SURGE	Most of the board is new. Headcount just unlocked.
NET_NEW_ROLES	Diffed against the previous crawl, not against posting dates.

Every signal carries a severity, a confidence level, the reason it matters, and the exact postings that triggered it.

---

The hard part is shutting up

A signal that fires on every company is not a signal.

The first calibration returned at least one signal for every single company tested, and AI_CAPABILITY_BUILD fired on 77% of boards — which in 2026 is just a description of the technology industry. That was treated as a bug, not a feature.

The test suite contains a deliberately boring control company that is required to produce zero signals. If it ever produces one, the build is wrong.

The pricing follows from the same idea: $0.08 per company that produced a signal, and nothing for one that didn't. If silence earned money there would be an incentive to loosen thresholds and bury you in noise.

---

Why the data is trustworthy

It reads the public, documented job-board APIs those seven platforms publish so companies can embed listings on their own careers pages. Third-party use is the intended use. No browser automation, no fingerprint evasion, no proxy rotation, no anti-bot arms race to eventually lose.

Two known-hard cases are handled explicitly rather than papered over:

Workday serves newest-first. Read the first thousand roles of a two-thousand-role board and you have read the fresh half — then reported "100% of this board opened in the last month". Truncated boards now report the true total, how much was actually read, and return null for freshness rather than a confident falsehood.

SmartRecruiters returns 200 OK with zero results for a company that does not exist, where every other platform returns 404. Counting that as a hit would attach an empty board to any string at all.

---

Actor on Apify Store →