MLOps monitoring,
made legible.

• 01 · Overview

Overview.

Designed a unified monitoring surface for an ML platform team running 40+ production models. The role: principal product designer working with platform engineering and SRE leadership.

• 02 · Challenge

The challenge.

Ops engineers were debugging in three places — Datadog for infra, MLFlow for experiments, a custom dashboard for drift — and stitching context manually after every page. P1 incidents averaged 38 minutes from alert to root cause.

• 03 · Process

Process.

Started with shadowing on-call rotations to map the actual debugging path. Ran 12 user interviews. Built information-architecture sketches against three real incident retrospectives, then prototyped at three fidelities — flow, mid-fi, hi-fi tokens.

• 12 stakeholder + on-call interviews
• Three retrospective walk-throughs replayed in prototypes
• Token-first design system aligned with platform engineering

• 04 · Solution

Solution.

A single canvas with three switching contexts: model health, deployment, drift. Status grammar borrowed from incident response (P1/P2/P3) so SRE and ML engineers shared a vocabulary. Replay scrubs through any incident in under 10 seconds.

• Unified canvas — one screen, three contexts
• Drift detector with rule-based + statistical thresholds
• Incident replay — scrub the timeline 60min around an alert
• Tokens shipped to engineering as a Tailwind preset