Outreach on fixed rules
Two founders and more than 6.000 schools and universities on the list. From March to November 2026 the system found the right contact, sent, followed up and sorted the replies. AI proposed. Fixed rules decided. A person took every warm reply. It's paused for now.
- Business
- StudyfinderAI, co-founded
- My role
- Designed and built it
- Ran
- March to November 2026
- Status
- Paused
emails sent over 85 send days
What it is: first emails and follow-ups to schools and universities, March to November 2026.
institutions contacted
What it is: schools and universities that got at least one email.
Why it matters
the system found the right contact on each institution's own website first.
Step 02, Contact researchcountries
What it is: the countries the institutions are in.
Why it matters
each country has its own holiday calendar. While a school is closed, nothing goes out there.
Step 03, Holiday checkemails a weekday, hard limit
What it is: the most the system may send in one day.
Why it matters
a hard limit protects the sending domain. Sending also pauses by itself when bounces pass 5%.
Step 05, Send gatesCounted from the system's own database, 23 March to 9 November 2026. I don't show reply rates. The count isn't clean enough to publish.
How it ran
Eight steps, every weekday it ran
Tap a step for the detail. The shape of each marker shows who is in charge.
The crew
Three agents, one job each
I called the system Project Nebuchadnezzar and named the agents after the crew of that ship in The Matrix. Each one has a written boundary: what it owns, and what it must never touch.
- Morpheus, the orchestrator. Reads the state of everything, decides the next job, starts it and keeps watching it. It talks to me in one message thread on my phone: a digest at the end of the day, and exceptions only.
- Link, the researcher. Imports institution lists, fills gaps such as website and city, finds and scores contacts, and prepares the holiday calendar per country. It never sends an email.
- Neo, the mailer. The only agent that touches the mailbox. It checks the gates, sends, plans the follow-ups and sorts the replies.
- A control room to watch them. A small game-style world where the agents walk around, a notice board shows the open tickets and a dashboard shows backlog, blockers and how fresh the data is. It only reads: nothing in it can send or change data.
351 commits from March to November 2026. The language model runs in the cloud first, with a local model on the always-on computer as backup. It never decides whether to send.
Step by step
What happens at each step
Guardrails
What keeps it safe
- AI never presses send. It researches, labels and summarises. Fixed rules decide what goes out.
- Automatic brake. Sending pauses by itself when bounces pass 5%, to protect the sending domain.
- Test and live are separate. Two databases. The system refuses to run when they don't match.
- Dry runs and backups. A dry-run mode, and a database backup before every automatic contact update.
- New subject lines need my yes. AI may suggest them. I approve them.
Built with
Stack, in plain words
- Three Python agents: research, mailing and an orchestrator.
- A small SQLite database, the Gmail API and a contact finder service.
- A language model for advice and labels, with a local model as fallback.
- An always-on computer with scheduled jobs and phone alerts.
- About 880 automated tests that run on every change.
What went wrong
What broke, and what I changed
- Auto-replies counted as real replies. Out-of-office mails and bounces were logged as replies, and some were even handed on as warm leads. I rebuilt the sorting to spot them in any language. It's also why I don't publish reply rates.
- School holidays shut out universities. The holiday check used school calendars for universities too, so German and Austrian universities got nothing for three weeks. I reopened them. A closure has to know what kind of institution it applies to.
- Old templates went out. 37 cold emails to schools used the old text. I set that lane to zero sends a day until the new copy was approved.
For your business
What you can take from it
- Let AI propose, let rules decide.
Use AI where judgement is cheap to check: research, labels, summaries. Keep the send button, the price and the promise in fixed rules.
- Build the brake before the engine.
Hard limits and automatic pauses protect what is hard to win back, like a sending domain or a customer list.
- People see only the exceptions.
A weekly report and alerts for the unusual. Your team spends its time on the replies that matter.
Sources: the system's configuration, decision log and database, read 2026-11-16. The numbers describe my engineering work, not StudyfinderAI product results.
Next caseStudyfinderAI websiteContact
Tell me what is stuck
Book 30 minutes on video, or write me a few lines about which AI tools your team uses and where it gets stuck.
consulting@mario-behrends.com