Skip to content

Emergency? Always call 911.

SafeAIOne
Trust Center

Performance report · Measured 2026-11-09 · Apache JMeter 5.6.3

How fast a report reaches dispatch, measured.

An incident is submitted; it has to become a record an agency can act on. This report measures that path end to end under real concurrent load, not a simulation and not a single hand-timed request. Every figure below comes from a recorded Apache JMeter run, and the original JMeter dashboards are published alongside it.

  • 170msSubmission to record visible in the dashboard API, median
  • 299msSame measurement, 95th percentile
  • 0%Error rate across 1,200 submissions in 3 runs
  • 9.8/sSustained completed throughput

1. End-to-end latency at the operating rate

500 submissions at 10 per second, sustained for 50 seconds, on the application’s default configuration. Held long enough to mean something, with a sample large enough to support percentiles.

10 submissions/second sustained for 50 s — 500 samples · 500 of 500 accepted · 0 errors
MeasurementSamplesBestMedianp90p95p99Worst
Record visible in the dashboard APIsubmission start → completed API response containing that exact record50013 ms170 ms280 ms299 ms347 ms493 ms
Submission HTTP responsewhat the submitting app waits for before it can confirm50095 ms791 ms1,492 ms1,684 ms2,821 ms5,135 ms
Dashboard list APIthe authenticated records call the dashboard itself makes19411 ms25 ms42 ms46 ms90 ms264 ms

The record becomes queryable before the submitting app gets its reply — the row is written early in the request, while duplicate detection runs before the response is sent. That is why visibility (170 ms median) beats the submission response (791 ms median). For staff watching the dashboard, the data is there first.

What “visible” means here

A completed API response containing the exact record identifier — measured, not inferred. It is reported as the upper bound of a bracket, because a poll can only prove the record was present somewhere between the moment the request left and the moment the response arrived. The lower bound for this run is 138 ms median. It does not include browser rendering: the dashboard refreshes on its own interval, so add up to one refresh period for on-screen appearance.

2. At the operating rate, and above it

Three runs on the default configuration: the operating rate measured twice, and the same system held at two and a half times that rate. No request failed in any of them.

Published runs · local production build · MariaDB 11.4
Offered loadAcceptedErrorsCompletedSubmit medianSubmit p95Visible medianVisible p95
10/s for 50 s500 samples500/50009.8/s791 ms1.7 s170 ms299 ms
10/s for 20 s200 samples200/20009.9/s378 ms1.1 s177 ms298 ms
25/s for 20 s500 samples500/500017/s3.3 s11.7 s1.7 s5.1 s

“Completed” is throughput actually achieved, counting only records confirmed to exist — a fast rejection is never counted as a processed report. At the operating rate the system keeps pace exactly and answers in well under a second. Pushed to two and a half times that rate it still accepts and stores every single submission, and records remain visible to the dashboard in under two seconds; responses to the submitting client simply take longer. Nothing is dropped.

3. The original Apache JMeter dashboards

Generated by Apache JMeter from the raw samples and published unmodified, with its own charts and per-request detail. These are the primary evidence for the tables above.

Percentiles in JMeter’s own tables differ slightly from the tables above — for the run in section 1, JMeter reports a 99th percentile of 3,399 ms where this page reports 2,821 ms. Same samples, different convention: JMeter interpolates between neighbouring samples, while the figures here use the nearest actual sample. Neither is adjusted; both are shown so the difference is visible rather than hidden. JMeter’s dashboards also cover only the submission request — the end-to-end visibility measurement is this page’s, taken by a separate observer.

4. Live site, sequential samples

Separate from everything above, and the only measurement here that includes TLS and real wide-area network latency: a handful of sequential requests against safeai.sequorademos.com. It says what one user feels when nothing else is happening. The sample is far too small for percentiles.

Sequential samples over the public internet
OperationSamplesResponse timesMeanErrors
Public homepage3841, 335, 420 ms532 ms0
Mobile incident submission2751, 904 ms828 ms0
Dashboard records API (authenticated)3943, 643, 723 ms770 ms0
CallAssist records API (authenticated)3686, 170, 157 ms338 ms0

One record (CR-261109-310) was submitted and followed through to the dashboard API: the submission returned in 737 ms and the record appeared in a completed dashboard API response 2.4 s after the submission began.

5. How we measured, and the limits

How it was measured

  • Load tool: Apache JMeter 5.6.3, against the production build of the app with MariaDB 11.4, on a workstation with 8 cores and 16 GB of memory.
  • Each request was a labelled test incident with no photos or contact details.
  • The database held 30 to 530 active records during the runs.
  • A separate monitor polled the dashboard and timed when each new record first appeared.
  • Only records confirmed in the database were counted. All test records were deleted afterwards.

Limits

  • Sections 1 to 3 ran on one machine shared by the app, the database and the load generator. Production servers will perform differently.
  • No CDN, proxy, TLS or internet connection, except in section 4.
  • “Visible” means the dashboard returned the record. Browser rendering was not timed.
  • No photos or videos were uploaded; those submissions take longer.
  • Phone (voicemail) intake was not measured.
  • These are measurements, not a service level guarantee.

Apache JMeter is a trademark of The Apache Software Foundation, which has not reviewed or endorsed these results. The JMeter dashboards in section 3 are published unmodified.

All figures come straight from the recorded run data. Measured 2026-11-09. Back to SafeAIOne

YOUR AGENCY. YOUR WORKFLOW.

See SafeAIOne in action.

A guided walkthrough, shaped around your team. Tell us a little about your agency and we’ll take it from there.

AI-assisted intake Incident records CAD integration

By submitting, you agree that our team may contact you about your request. See our Privacy Policy. Please do not include sensitive incident information.