//benchmarks

Measured results, with the method attached

Before-and-after figures from delivered engagements. Each entry states the workload it was measured on and how it was measured, because a performance number without its method is marketing, not evidence.

Client names are withheld under NDA. Figures are from production systems, measured on identical hardware before and after unless stated otherwise. They describe what those systems achieved under those workloads — not a guarantee of what a different system will achieve. We will happily walk through the raw test output for any of them on a call.

p99 API latency

840 ms61 ms

B2B order API, 1.2k RPS peak

method

Query plan rewrite, composite indexes, response caching with tag invalidation. Measured with k6, 20-minute soak, 4 vCPU app nodes.

Sustained throughput

620 RPS9 400 RPS

Read-heavy catalogue service

method

Connection pooling, read replicas, payload compaction and an async write path. Same hardware budget; measured on identical instance sizes before and after.

Monthly infrastructure cost

€4 180€1 560

Kubernetes platform, 3 environments

method

Right-sized requests and limits from 30 days of usage data, spot node pools for stateless workloads, storage-class review. Cost reported by the provider, not estimated.

Deployment lead time

4.5 hours11 minutes

Monolith with manual release steps

method

Pipeline parallelisation, test-suite partitioning, automated migration gate and blue/green cutover. Median over 60 consecutive production deploys.

Mean time to recovery

3 h 20 min18 min

Payment integration platform

method

Structured logging with trace correlation, actionable alert policy, documented runbooks and quarterly restore drills. Median across 12 months of incidents.

5xx rate under peak

2.4 %0.02 %

Seasonal retail traffic, 8× baseline

method

Backpressure and queue-based load levelling, idempotent retries, graceful degradation of non-critical reads. Measured across a full peak week.

//verification

How we would prove it on your system

The same procedure every performance engagement starts with. It is deliberately boring and entirely reproducible.

  1. 01

    Capture a representative traffic profile from production — endpoint mix, payload sizes, concurrency, cache state.

  2. 02

    Build a k6 or JMeter harness that replays it against a staging environment sized like production, and record the baseline.

  3. 03

    Profile under that load. Fix what the profile shows, in order of measured cost, one change per commit.

  4. 04

    Re-run the identical harness. Publish both runs, including the changes that did not help.