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
Evidence before narrative
The correlation engine creates the incident record. AI summarizes the evidence rather than inventing a conclusion.
Separate stores for separate work
PostgreSQL carries operational state, ClickHouse serves telemetry analysis, and Redis supports low-latency coordination.
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.