agentificial
case study · anonymised
·Infrastructure· Mar – Oct 2026

Self-hosted outbound email at fleet scale, and the six ways it failed silently

We built and operate cold-email infrastructure of about 160 domains and 790 mailboxes on self-hosted mail servers. Most of the work since has been making silent failures visible.

~160 / ~790
domains / mailboxes in the fleet
105,917
messages in 14 days of DMARC aggregate reports
0.00–0.32%
fleet reply rate: infrastructure ready, audience is the limit

The problem

The group wanted cold-outbound capacity at scale without paying per-seat SaaS pricing. Once volume existed, the harder question was why deliverability and replies stayed poor.

Mail infrastructure rarely breaks loudly. A closed port, a misrouted relay or a debug sink can look exactly like a quiet week.

What we built

  • A mail server stack on a cloud VPS (Mailcow) with two egress relays running Postfix, linked over WireGuard, one IP per relay.
  • Per-domain DKIM signing with rspamd, SPF and DMARC, and the group's own tracking domains behind Caddy.
  • A campaign engine on Celery and PostgreSQL, with warm-up through Instantly.
  • Monitoring: DMARC aggregate parsing (parsedmarc), Prometheus and Grafana, a KPI watchdog, SMTP bounce ingestion, and nightly backups.

Result

  • In the 14 days to 2 October 2026, DMARC aggregate reports covered 105,917 messages: 77,676 to Microsoft (100% DMARC pass, 12 quarantined) and 24,901 to Google (100% pass).
  • Failure 1: a firewall rule dropped port 587 for nine days and warm-up scores decayed to zero.
  • Failure 2: a 99.8% volume collapse went unnoticed because the alarm only fired at exactly zero. Fixed on 5 September 2026 with a rolling 24-hour baseline, backtested with zero false positives.
  • Failure 3: a debug SMTP sink silently black-holed 175,828 messages over 3.5 days while every layer reported success.
  • Failure 4: the deliverability scorer was inert, with all 5,456 mailbox rows pinned at 72.0. Fixed on 23 September 2026 with cohort scoring, bounce ingestion and a KPI watchdog.
  • Failure 5: 460 mailbox passwords went stale after a server rebuild.
  • Failure 6: a relay had been null-routed since 15 July 2026 while a second brand's sending accounts still pointed at it. The only signal was "sent: 0".
  • The principle that came out of it: zero bounces plus zero opens plus zero replies is an outage signature, not a deliverability win.

Scope and limits

  • The results are weak and we say so. Fleet reply rate is 0.00–0.32% (0.03–0.23% per campaign) and Gmail inbox placement measured roughly 0.6–6%.
  • In early September 2026 the platform had capacity for about 57,000 messages a day and sent 44 real emails in 19 days, because no active campaign had a safe audience. The infrastructure is ready; audience and inbox placement are the limit.
  • No appointments or revenue are attributed to email outbound.
  • The servers sit in the client's own cloud account. We operate them; we do not own them.
· Could be you next

Same problem,
different stack.

Tell us the workflow that should already be done by the time your team logs in. We'll come back with a one-page assessment.