Job Outreach
Job letters from my own mail server, measured without a pixel
CRM Campaign shows a bank campaign on simulated data; here the same craft meets real recipients. I am looking for a CRM job, so I send my own applications as a small CRM campaign: AI agents found seven Czech companies that invite a CV by e-mail and drafted one letter for each, and Mautic, an open-source marketing tool, sends them through my own mail server, so they pass spam filters as authenticated mail (SPF, DKIM, DMARC). There is no tracking pixel: as I read EU law it would need the reader's consent, and mail apps open images by themselves anyway. Clicks go through my own redirect, which every letter discloses. The first proof came on 11 November 2026: one letter went through a company's own mail gateway and Microsoft 365 to a person, who answered within 46 minutes. It was a decline: 1 reply out of 5 letters, with 0 human clicks (REAL DATA, n = 5). It shows that the chain works on real people; it says nothing about what makes a letter work.
After the simulated CRM Campaign, the same craft on real recipients: my own job applications, one letter per company, sent through Mautic and my own mail server. Delivery, replies and disclosed clicks are counted; opens are not.
How it works
The idea: CRM Campaign runs the craft on simulated data for a fictional bank. This project tries the same craft on real recipients, with my own job search as the campaign. It is a source of measurements first and a way to replies second, and it measures only what can be measured without breaking anything.
The contacts: AI agents searched Czech careers pages for an address the company itself gives for sending a CV. Seven companies passed, each checked on its own page: no guessed addresses and no personal addresses of employees. This page names none of them.
The letters: one Czech letter per company, with a sentence or two about that company from its own page, a link to a case study, the CV attached, the note that AI agents write my code, and a way out: a short reply, and no more mail follows. Each letter has a plain-text part and a designed HTML part with no images.
Sending: a second Mautic on my PC, separate from the CRM Campaign demo, holds the contacts and the letters. It sends through my own mail server over SMTP with TLS, and the server signs each letter with DKIM. A send script refuses a letter unless it was approved in its current wording, the address is the verified one, the contact never got mail before and fewer than 10 company letters left that day. Letters go one at a time, and the server log is read after each.
No pixel: Mautic's open pixel, its link redirects and its unsubscribe header are switched off; from a PC they would point at 127.0.0.1 anyway. Before a batch, a proof run sends every letter to a local mail catcher and checks the raw message: no pixel, image or remote resource, no trace of the PC, links exactly as in the letter, the plain text identical to the letter, every sentence in the HTML, the CV byte for byte and no phone number. The checks were shown to fail on a copy with a pixel and an unsubscribe header added.
Clicks, disclosed: every link is go.aspandos.com plus a random code per letter and link, and the letter says so in one line. The server answers with a plain redirect to the normal page and logs the time, the code and a short user agent: no IP address, no cookie, nothing added to the page's address. Mail security scanners open links on their own, so a click counts as human only if it comes more than 60 seconds after delivery from a user agent that is not a known scanner.
Real use, batch 1 (REAL DATA, n = 5, as of 11 November 2026): five letters went out on 11 November after my yes, and the receiving servers accepted 5 of 5 (SMTP 250, which means accepted, not delivered to an inbox). Inbox placement is proven in three places only: my own Gmail copy of the first letter (Primary tab), a test mail from the same server to a university Microsoft 365 mailbox (Focused tab), and company A.
Company A is a media company with its own mail gateway in front of Microsoft 365. A person there answered 46 minutes after delivery, and the answer was a decline. So for one real recipient the whole chain worked from end to end: sent, accepted, through the company's gateway and Microsoft 365, read by a person and answered. It shows nothing about the other four letters, and nothing about interest in the case study: the gateway's link scanner opened all three links in the second of delivery, so the raw count is 3 clicks and the human count is 0. No other letter has a reply yet; replies are counted for 14 days, and two more letters wait for my yes.
What it is not: CRM work at a company. It is Mautic for my own job search, seven contacts, and not production experience with Mautic or any commercial CRM tool. Five letters can show that the chain works from end to end; they cannot show what makes a letter work, and this page does not try.
Limits and decisions
- Five letters, no conclusions. n = 5 is a count, not a sample. The page reports counts, never rates, and says nothing about which letter, subject line or kind of company works better. Batch 1's 14-day reply window runs until 25 November 2026, so the figures can still change.
- Inbox placement is proven only three times. SMTP 250 means that a receiving server accepted the letter; it says nothing about the inbox, the spam folder or a quarantine. Two of the three proofs are test mailboxes of my own, so for four of the five companies the placement is unknown.
- Clicks are a heuristic. The 60-second rule and the list of scanner user agents remove the scanners known so far. A scanner that waits longer or poses as a current browser would count as a person, and a person who reads the letter without clicking counts as nothing.
- No opens, on purpose. A tracking pixel reads information from the recipient's device, which in the EU needs consent: my reading of the ePrivacy rule as Czech law has it (Act No. 127/2005, section 89(3)), not checked by a lawyer. Opens are also unreliable, because mail apps fetch images by themselves. So opens are not counted, and every letter says so.
- One reply arrived 16 minutes late. My server's spam scan ran over its 8-second limit on company A's reply and asked the sender's server to try again later, which it did 16 minutes on. Nothing was lost, and the limit is now 30 seconds. The 46 minutes above count to the time the reply was sent.
- The letters are agent drafts. AI agents drafted the letters, and each batch went out after my yes. The sample in the stills was written for a fictional company and is not one of the letters that were sent.
What's inside
- The first proof on real people
- No tracking pixel.
- Disclosed clicks
- A scanner filter
- Send guards
- My own mail server
Built with: Mautic, Mail server, Python, Claude Code
Previous project: AI Toolbox · Next project: AudioVJ · All projects