AI Observability

A module that brings your AI infrastructure — LLM services, GPUs, and agents — into a single view of health, cost, and performance.

AI Observability

AI Observability extends SAF to full observability of your AI infrastructure — its health, cost, and reliability. It brings together telemetry from LLM services, GPUs, AI agents, and local AI clients into a single view of status, spend, and performance, with ready-made scenarios for operations, ML, and FinOps teams.

Who it's for:

ML/AI platform engineers and DevOps
FinOps
CTOs and AI leaders.

key features

Single view of AI infrastructure and GPU health

A consolidated operational picture of your entire AI landscape, with one-click drill-down to GPU health:

live status of LLM models and AI services: requests, traces, events, tokens, and cost.
GPU status: utilization, video memory, temperature, and hardware errors.

LLM Costs

Control over LLM request consumption and cost by model and provider:

tokens and cost broken down by model, provider, and cost trends.
detection of abnormal cost growth and your most expensive requests.
side-by-side comparison of model efficiency by consumption.

AI agent and trace analysis

A breakdown of AI agent behavior and request structure down to the individual span:

requests, responses, tool calls, tokens, and latency across conversations.
drill-down to traces: span details, a request map, and errors.
pinpointing bottlenecks and root causes of failures at the tool-call level.

Monitoring of local AI clients

Observability for developers' AI tools, shown here with Claude Code and Codex:

conversation activity, token consumption, application events, and errors.
usage broken down by user and project.

Monitoring of local AI clients

Observability for developers' AI tools, shown here with Claude Code and Codex:

conversation activity, token consumption, application events, and errors.
usage broken down by user and project.

AI infrastructure health in Service Monitor Toolkit

6 ready-made global metrics that bring the AI landscape into the resource-and-service health model:

GPU (temperature, utilization, memory), request duration, trace errors, and service availability.
early degradation signals before your users notice them. - configurable severity thresholds and a link between technical state and service impact.