Who to contact today, and why now.
A daily list of the 10 B2B SaaS accounts worth contacting, built every morning from public job boards, funding news and pricing pages. Detection and scoring are plain code you can read.
Timing is most of outbound.
The same email lands differently the week a company hires its first Head of RevOps, raises a round or changes its pricing. Those moments are public, but scattered across job boards, news and company websites, and nobody checks 200 of them every morning by hand.
So this checks them automatically, scores what changed, and hands over a short list with the reason attached, so the first line of an email can be about something real.
Sources, detection, scoring, digest.
- 01
Sources
Public job-board APIs (Greenhouse, Lever, Ashby), five funding-news RSS feeds, and each company's pricing and careers page.
- 02
Detection
Plain code compares today with yesterday: new job IDs, new headlines naming a company and a round, page text that really changed.
- 03
Scoring
Each signal type has a weight that halves every 7 days. Points add up per company, and a template writes the one-line reason.
- 04
Digest
The top 10 goes to a private daily digest with draft first lines, and to the public dashboard without them.
Every morning a GitHub Actions job collects, detects, scores and commits the results to the repository as files. Raw facts are written once and never edited; signals are recomputed from them, so a rule fix can be replayed over the whole history.
Every company proven, not assumed.
408 well-known B2B SaaS companies went in. A company stayed only if one of its job boards answered with open jobs and the board was shown to be that company's: Greenhouse's company name, or the company's own name or domain in its Lever or Ashby job descriptions. 257 passed, and 200 were picked across five groups (sales and marketing tools, developer tools, data and AI, security, finance and operations).
Six boards matched by name but belonged to a different company. Reading each one's own "About us" text caught them:
- CubeThe board was CUBE, a RegTech company, not cube.dev.
- Neon (twice)One board was a Brazilian fintech, the next a game-payments company. Neither is the Postgres company.
- MosaicA private-markets deal-modeling tool, not mosaic.tech.
- RunwayRunway the business-planning tool, not RunwayML.
- RemoteGreenhouse's company name was "General Assembly Remote Jobs". The check now needs the name to start with the company's name, not just contain it.
Pricing and careers pages were then found for each company: 152 pricing pages and 170 careers pages are watched (322 in all). The rest are blocked by robots.txt, need JavaScript to render, redirect off-site, or don't exist.
Weights in a config file.
| Signal | Points | Rule |
|---|---|---|
| Funding round | 40 | Several articles about one round count once |
| Sales or marketing leader hired | 30 | VP, Head or Director of Sales, Marketing, RevOps; CRO, CMO |
| RevOps / sales ops / GTM engineer role | 25 | Up to 2 per company; one title in several locations counts once |
| Sales hiring burst | 20 | 3 or more new sales-rep roles in 7 days |
| Pricing page changed | 15 | A price changed, or more than 5% of the visible text |
| Careers page changed | 5 | Same rule as pricing |
| Any other new job | 1 each | At most 5 points |
Weights assume a seller of RevOps and sales-automation services, so GTM hiring counts most. Each signal's points halve every 7 days and drop out after 14. An account listed in the last week is skipped unless something worth 25 points or more is new. Every account's line says exactly which signals put it there, with links to the job post or article.
What broke, and how I fixed it.
Funding matched stories that weren't rounds.
A full test run, read signal by signal before going live, matched a merger ("Ada to merge with Dialpad … $600 million"), a government-policy story, and "1Password tops $500-million ARR". Merger, acquisition and IPO wording now excludes a headline unless it names a round, and a dollar amount alone no longer counts.
Job titles landed in the wrong family.
"Industry Principal, Office of the CMO" counted as a CMO hire, and an R&D "business systems" engineer as RevOps. C-level titles must now start the title, and only sales or revenue systems roles count as RevOps. Each fix has a self-test case.
One job in two cities counted twice.
"Product Marketing Operations Manager | Germany" and "| Ireland" were scored as two hires. Titles are now compared without the location.
A second run on the same day would have overwritten the first.
The first run happened just after midnight UTC, so that morning's scheduled run had the same date. Each date is now collected once; a second run skips instead of overwriting saved data.
The dashboard's funding filter showed nothing.
The recent-signals table kept the newest 300 rows, and 1,208 backfilled jobs pushed the funding rows out. It now keeps every non-job signal plus the newest 300 jobs.
Collection is free; the model is a rounding error.
Job boards, feeds and pages are public and need no keys. The first daily run read 200 of 200 boards, all 5 feeds and 321 of 322 pages in 234 seconds of GitHub Actions time, on the free plan.
Claude Haiku 5.5 is used only where words are needed: draft first lines for the top 10, and a plain-language summary when a pricing page changes. A one-account test call cost $0.00012. A hard cap of $0.05 per run is enforced in code before every call, and every run's cost is logged and shown on the dashboard.
Polite by construction.
Every request names itself as a research bot that fetches each page at most once a day. robots.txt is checked before every page and every redirect hop; requests to one site are at least a second apart. Sources that refuse the bot, with a 403, a 429 or a robots.txt rule, are left out, never worked around; several candidate news feeds were dropped that way.
No personal data is collected. Draft first lines are addressed to a role, never a named person, go only to a private daily digest, and are never sent to anyone.
False-positive review: coming after week 1.
After seven days of runs, a seeded random sample of 20 signals per type (jobs, funding, page changes) goes to a human reviewer with the evidence link for each. The false-positive rate per type, with a confidence interval, will be published here and on the dashboard. Until then, no accuracy number is claimed.
What it's built with.
- GitHub Actionsruns every morning, commits the day's data as files; no database
- TypeScript / Nodecollectors, robots.txt parser, detection rules, scoring, self-tests
- Claude API (Haiku 5.5)draft first lines and pricing-change summaries only; capped per run in code
- Static dashboard on Vercelrebuilt from the committed data on every daily commit