RAMIL.ORG Independent systems building and technical operations
← All systems

AI-powered email infrastructure observability

MX Sentinel

A production operations layer that helps mail operators understand why delivery is failing — before the incident becomes a support queue.

Problem

Email failures are noisy. Operators were forced to assemble a story from mail logs, DNS records, aggregate reports, and customer tickets while delivery was already degraded.

Constraints

  • High-volume telemetry with privacy boundaries
  • External DNS and provider signals can be incomplete
  • A diagnosis must be explainable enough to act on

Approach

A Go-based event system collects metadata, validates mail configuration, and correlates changes and delivery signals into operational incidents. An AI layer writes a readable hypothesis and remediation path from that evidence; it is not the source of truth.

Architecture

Key decisions

01

Evidence before narrative

The correlation engine creates the incident record. AI summarizes the evidence rather than inventing a conclusion.

02

Separate stores for separate work

PostgreSQL carries operational state, ClickHouse serves telemetry analysis, and Redis supports low-latency coordination.

03

Metadata-only diagnostics

The system avoids email bodies and hashes sensitive recipient data.

Failure handling

Collectors spool safely when a bus is unavailable. Jobs are idempotent, retries are visible, and operators can distinguish an infrastructure outage from missing evidence.

Result

A single operational surface for moving from delivery symptom to likely cause, with the signals preserved for review.

What I’d change next

Add richer long-term baselines and more provider-specific feedback loops.