AI
The 6-8 Week FinOps Problem
The standard FinOps review cycle takes 6 to 8 weeks: identify what’s wasting money, deploy a fix, wait for the next cycle to see results, then repeat. That pace made sense before daily deployments. Autoscalers help, but they act only on current data. No open-source tool combines observed metrics with forecasted values to drive proactive cost decisions. The tools already exist: Prometheus, Thanos, OpenCost, Karpenter. The problem is that nothing integrates them.
Auger: Three CRDs and One kubectl Apply
Auger ships three CRDs. The ForecastPipeline defines cadence, metric source (Prometheus, OpenCost, or promxy), and which model to use. The ForecastModel abstracts serving via the v2 Open Inference Protocol, so any compliant framework connects. Toto is pre-trained on close to a trillion data points and needs only historical context (7 days, a month) to generate forecasts. ARMA is a classical univariate model that supports the v2 protocol natively. One kubectl apply wires all three stages together.
Standard Formats and Multi-Model Inference
Auger defines a single metric frame format. Any source (Prometheus, OpenTelemetry, OpenCost) converts to it with one Python class. Models output a forecast envelope: mean value plus confidence bounds. Adding a new model takes two Python functions. Custom recommendation policies take one Python class. Golang runs the Kubernetes operator and reconciliation loop. Python runs inference. Contributors only touch Python. The operator layer stays fixed across use cases, and you can run multiple models against the same data to compare results before picking one.
Drift Detection and Human Approval
Auger compares forecasted values against actual cluster state when the prediction window closes. If accuracy drops below threshold, it raises a Kubernetes event. State of FinOps 2026 data: more than 50% of teams still set resource limits and requests manually. Recommendations appear as a Recommendation CR with projected savings, peak CPU, and peak memory. A human reviews and patches the spec with an approval status before Auger applies anything. Model drift and hallucinations are real risks. The approval gate prevents automated changes from propagating errors.
Demo: Toto and ARMA on a Kind Cluster
The demo runs on a local kind cluster with two worker nodes, an Auger controller manager, and separate ARMA and Toto serving deployments. OpenCost provides cost metrics. A Grafana panel shows mean forecast and two confidence bounds per node. ARMA’s line is flat because the cluster was one or two days old at demo time and ARMA needs more historical data. Accuracy is measured by training on the first 70% of available data and predicting the remaining 30%. One Recommendation CR targets a specific worker node for removal.
Notable Quotes
generating recommendations, it has to be deterministic because we can’t rely always on nondeterministic AI models. Satyam Bhardwaj · ▶ 7:43
as per state of phops 2026 report we still have more than 50% of the people setting uh limits and requests manually. So nobody relies on AI to that particular extent. Satyam Bhardwaj · ▶ 14:35
this is a pre-trained model uh close to trillion data points. So I do not have to train that on my matrix from my cluster. Satyam Bhardwaj · ▶ 19:42
Key Takeaways
- The FinOps review cycle takes 6 to 8 weeks; Auger tightens it with continuous forecasting.
- Auger’s Go/Python split means ML contributors only write Python; the Kubernetes operator stays fixed.
- State of FinOps 2026 shows 50%+ of teams set limits manually, so human approval gates every recommendation.
About the Speakers
Ankur Singh is a Senior Site Reliability Engineer at Red Hat, where he builds and manages Azure Red Hat OpenShift (ARO), handles observability, and responds to customer incidents.
Satyam Bhardwaj is a Software Engineer in Mirantis’ Open Source Program Office. He advocates for open-source adoption and contributes to multicluster management projects including Cluster API, cod, and kzeros.