All DevOps case studies / Case study 01 / 05
Vantrel Pay
On-call copilot (agentic SRE)
An on-call copilot that has already investigated when the engineer wakes up
In short
When an alert fires at 3 a.m., AI agents query the metrics, the latest deployments and past incidents, then post the likely cause and a ready-to-copy fix command in Slack.
- Client
- Vantrel Pay
- Sector
- Online payments
- Size
- 450 staff, 2 million transactions a day
- −70%time to find an incident’s cause
- 2 minto a first diagnosis, instead of 20
- 100%of fixes approved by a person
The context
At Vantrel Pay, every minute of downtime blocks payments. When an alert wakes the on-call engineer, the first twenty minutes go into opening dashboards, digging through logs and looking for what was deployed recently.
The challenge
- 01
The useful information is scattered across Prometheus, GitHub, runbooks and old Jira tickets.
- 02
At night, one engineer has to piece it all together, under pressure.
- 03
An AI must never change production without approval.
- 04
The diagnosis must be checkable: every claim, its evidence.
Our approach
An orchestrator (LangGraph) fires on every Alertmanager alert. It sends three agents in parallel: one queries Prometheus to locate the spike, one lists recent pull requests and deployments on GitHub, the third searches runbooks and past incidents indexed in Qdrant. The orchestrator cross-checks the findings and posts a Slack summary with the likely cause, the evidence and a fix command or Terraform script that a person applies in one click.
The solution, step by step
- 01
The alert fires
03:12: latency on the payment API shoots up.
- 02
The orchestrator dispatches
Three agents set off in parallel: metrics, changes, memory.
- 03
Evidence gathered
A connection spike, a PR deployed six minutes earlier, an identical incident last year.
- 04
A summary in Slack
Likely cause, evidence and a fix command, ready to copy.
- 05
A person approves
The engineer applies the fix: latency is back to normal.
Results
- −70%average diagnosis time
- −45%average incident duration
- 2 minto the first Slack summary
- 0production actions without approval
What we deliver
- Orchestrator and specialised agents (LangGraph)
- Alertmanager, Prometheus, GitHub and Jira connectors
- Index of runbooks and past incidents (Qdrant)
- Incident summaries and fix commands in Slack
- Audit log of every diagnosis
Technologies
- LangGraph
- Python
- Prometheus
- Grafana
- Qdrant
- Slack API
Timeline
8 to 10 weeks
Team
1 SRE, 1 AI engineer
Illustrative cases: clients and figures are fictional, representative of the solutions we deliver.