Model the end-to-end response time and service capacity of an AI code review pipeline. Combine queue delay, prompt preparation, model time, post-processing, and serial review passes, then compare the result with a target service-level time. The calculator also estimates reviews per worker-hour and the concurrency needed for a chosen arrival rate. Use measured percentile timings when planning an SLA; simple averages can hide peak-hour congestion.
Active service time = preparation + passes × (model time + post-processing)
End-to-end latency = queue delay + active service time
Concurrency = ceiling(arrivals per hour ÷ capacity per worker)
End-to-end latency is the estimated wait from submission to completed review. Positive SLA headroom means the estimate is below the entered target. Minimum concurrency covers the stated average peak rate without utilization safety margin.
For production planning, enter measured p95 or p99 component times and add utilization headroom; this deterministic estimate does not model queue variability.
With 4 seconds queued, 1.5 seconds preparation, and two passes of 18 model seconds plus 2 post-processing seconds, active time is 41.5 seconds and total latency is 45.5 seconds. A 60-second target leaves 14.5 seconds of headroom.
No. It is the mathematical minimum for the entered rate and active service time, so operational headroom should be added separately.
The calculator assumes one queue before a review workflow whose passes then run serially.
Yes, but adding several component percentiles is conservative and is not necessarily the exact end-to-end p95.
Required concurrency is zero, while latency and per-worker capacity still calculate.
Include network time in prompt preparation, model time, or post-processing according to where it is measured.
| Component | Applied | Unit |
|---|---|---|
| Queue delay | Once per review | seconds |
| Preparation | Once per review | seconds |
| Model + post-processing | Each pass | seconds |
| Arrival rate | Capacity comparison | reviews/hour |