Company: Cyvl
Industry: Software & Technology (GovTech)
Cyvl, a Somerville, Massachusetts startup, uses LiDAR and AI to inspect public infrastructure — measuring signs, cracks, and potholes so public works departments know where to spend taxpayer dollars. It’s onboarded more than 500 U.S. communities, processed over a petabyte of data, and helped one East Coast city reallocate $79 million in capital spending.
The challenge
Cyvl’s sensor rigs — mounted on cars, pushed along sidewalks, or dragged behind bikes — run LiDAR and imagery through more than 100 proprietary machine learning models, doing in about a week what used to take an army of student interns a summer to cover a fraction of a city.
Three years ago, each stage of that pipeline was owned by a different engineer, and logs were the only debugging tool available. Datasets were large enough that pulling logs to debug was slow, and the further an error traveled down the ML pipeline, the more expensive it got to catch — a real problem when a city council sometimes needed answers within a couple of hours.
As Cyvl won product-market fit, a new problem emerged: an event-driven, Kubernetes-based architecture that scaled sharply whenever the company processed a new city — sometimes several at once, like Atlanta, Nashville, and New York City together — then scaled back down. That created a high-cardinality problem big enough that the team went back to its Grafana account team and said “The bill’s too high. What can we do? Maybe we can’t work with Grafana anymore.”
The solution
Cyvl’s first move was consolidating engineering around one shared operating picture in Grafana Cloud, with Slack alerts and dashboards replacing single-owner log-diving. The team then built an AI agent framework on Grafana IRM: when an alert fires overnight, it lands in Slack and IRM with a dashboard link, and an agent – using the Grafana MCP server – investigates, proposes next steps, and can act directly in Cyvl’s production account. Engineers originally approved or rejected each action with a check or X emoji in the Slack thread; as confidence grew on common errors, the team loosened those guardrails and now pages a human only when a fix doesn’t hold.
On cost, Grafana’s team pointed Cyvl to Adaptive Metrics after the cardinality spike. It showed exactly which metrics fed no Prometheus alert or dashboard, so the team could turn them off with a click instead of assigning an engineer weeks of manual cardinality tuning. Cyvl has also started using Grafana Assistant to query its own Grafana Cloud data directly — one more way a small team covers more ground without adding headcount.
Cyvl’s stack today runs on that same foundation: Grafana Cloud Logs, Metrics, and Traces for the core telemetry, Grafana IRM for response, Grafana Cloud k6 for load testing, and Grafana Faro for frontend metrics — all instrumented with OpenTelemetry. Cyvl wrapped its OpenTelemetry setup into an internal library so engineers get metrics, traces, and logs configured in three or four lines, and treats observability like unit tests: code doesn’t ship without it. The Grafana MCP and a set of Claude Code skills also let non-engineers build their own dashboards by prompting in their own tools – product uses Faro to track feature usage, sales spots upsell signals, and operations tracks sensor GPS logs and data-processing timelines.
The impact
With costs under control and alerts triaged automatically, Cyvl’s team spends nights sleeping instead of debugging – and its non-engineers now build their own dashboards.
- Cut cloud computing costs by 42%, directly improving product margins
- Cut the Grafana Cloud metrics bill by 73% with Adaptive Metrics — in days, not weeks of manual tuning
- Cut overnight incident response time by letting an AI agent triage alerts as they arrive
- Extended Grafana dashboards beyond engineering to product, sales, and operations teams
- Standardized OpenTelemetry setup to three or four lines of code per service
- Onboarded 500+ U.S. communities and processed more than a petabyte of infrastructure data
- Helped one East Coast city save and then reallocate $79 million in taxpayer capital spending