Features
Continuous high-resolution profiling timelines
Operation duration and resource utilization tracking
LLM generation tracing with per-step timing
Token throughput and latency breakdowns
System-level metrics for CPU, GPU, accelerators
Error monitoring for device-level failures
Inference telemetry for AI agents (autodebug)
Low-overhead CUDA kernel attribution
Claude Code integration for AI debugging
Profiler CLI and Python API
Support for vLLM, SGLang, PyTorch, TensorRT-LLM
20+ out-of-box evals for RAG, agents, safety, security
Custom evaluators encoding domain expertise
Auto-tune evals from live feedback
Distill evals into Luna models for 96% cost reduction
Luna Studio for trustworthy evaluations at low cost
Eval Engineer integration with Claude and Codex
Insights engine identifying failure modes and prescribing fixes
Capture groundtruth from synthetic, dev, and production data
Subject matter expert annotations
Guardrail policies blocking harmful responses
Eval scores control agent actions, tool access, escalation paths
Low-latency evaluation on L4 GPUs
Ingest models, prompts, functions, context, datasets, traces, MCP server
Pre-production evals become production guardrails without glue code
Trace-based analysis with millions of signals per session
