Lumiere Analytics

A dashboard buckling under its own data

Lumiere's BI tool was the heart of their platform — finance teams at 200+ mid-market companies relied on it daily. But after raising their Series B and onboarding bigger customers, the cracks showed. P95 dashboard load times hit 28 seconds. Their Redshift bill had tripled in a year. And the engineering team was rewriting the same caching logic for every new feature.

The CTO put it plainly: "We can't grow the customer base because every new tenant slows everyone else down. We need to fix the foundation before we can ship anything new."

Re-architect the data plane, leave the product alone

We resisted the urge to rewrite everything. The React frontend was working — the problem was downstream. We zoomed in on the data layer:

• Migrated from Redshift to a ClickHouse + Postgres hybrid for hot/cold separation
• Built a query planner that pre-aggregates common dashboards on write
• Added a Redis-backed cache layer with TTL tied to dataset freshness
• Rewrote the API layer in Go to handle 10× concurrent queries on the same hardware
• Instrumented everything with OpenTelemetry → Grafana for ongoing visibility

We shipped it in 4 phases over 14 weeks, with zero customer-facing downtime.

Faster, cheaper, and ready for 10× growth

The numbers tell the story:

"Skill Horizon did what our last two contractors couldn't — they understood the data layer deeply enough to fix the actual problem. Three months in, we're shipping features again, customers stopped complaining about speed, and the infra bill is genuinely cheaper. They're our default partner now."

The toolkit