parelyxObservability
Evidence for every decision.
Optimisation claims are only as good as the evidence behind them. Parelyx records what it decided, why, and what happened — so you can verify the value on your own traffic.
Parelyx console
no deployment connectedNo inference traffic connected
Connect your inference traffic to inspect execution performance. Every chart in the console is built from your own requests — we don't show sample data.
01Measures that matter
What each measure tells you.
Clear definitions, so a number means the same thing to your engineers, your finance team and ours. Values only ever come from your own traffic.
Decision
- Decision record
- The strategy and endpoint selected for each request.
- Every outcome can be traced back to the decision that produced it.
- Decision overhead
- Time spent deciding, separate from execution time.
- Proves the optimisation layer is not the bottleneck.
Latency
- End-to-end latency
- Time from request received to response returned.
- The latency your users actually experience.
- p50 / p95 latency
- Median and 95th-percentile end-to-end latency over a window.
- Typical experience and the slow tail most users still see.
- p99 latency
- 99th-percentile end-to-end latency.
- Exposes rare but severe delays that averages hide.
- Time to first token
- Time until the first streamed token reaches the client.
- Perceived responsiveness in interactive and streaming interfaces.
- Time per output token
- Average generation time per token after the first.
- Separates generation speed from queueing and prompt processing.
Cost & effectiveness
- Token usage
- Input and output tokens per request, before and after optimisation.
- The primary driver of inference spend.
- Cost per request
- Estimated from token usage and a configured rate card. An estimate, not an invoice.
- Unit economics by workload, model and team.
- Optimisation outcome
- Whether reuse was used and how many tokens optimisation removed.
- Shows what each optimisation actually did.
- Net optimisation effectiveness
- Saving attributable to an optimisation minus its own cost, compared with a baseline.
- Tells you whether an optimisation is actually worth keeping.
Quality & health
- Quality evaluation
- Scores from automated evaluators and, where provided, human feedback.
- Ensures cost and latency gains are not paid for with quality.
- Provider health
- Error rates and latency by model provider and endpoint.
- Informs selection and highlights degraded providers.
02How to read it
Baseline first. Then decide.
- 01Connect with optimisations off
Establish latency, cost and token baselines for each workload on your real traffic.
- 02Enable one strategy at a time
Compare against the baseline so each change is attributable.
- 03Keep what creates net value
Disable anything that does not pay for itself — the engine is built to bypass it anyway.