Loading…
A FinOps and platform engineering reset that took AfriLogistics from a $4.1M monthly cloud bill to $1.55M without retiring a single feature.
“We were burning a Series C round on AWS and didn't realise it. Spalce came in, mapped where the money was actually going, and stopped the bleeding without us slowing down a single delivery.”
AfriLogistics had grown its AWS footprint from a startup posture to a 1,200-microservice fleet in eighteen months. Compute spend was outpacing revenue growth. We ran a five-month FinOps and platform engineering engagement: introducing a workload-aware cluster autoscaler, rightsizing 84% of services, moving the analytics tier to Spot + Graviton, and re-architecting the three services that were responsible for 41% of the bill. Monthly compute spend fell 62% and unit economics on parcel delivery improved by $0.34 per shipment.
AfriLogistics had scaled from twelve engineers to two hundred and thirty in eighteen months on the back of a Series C. Its AWS bill had grown faster: from $180k a month to $4.1M, and the curve was still bending up. The CFO's signal to the board had been that gross margin would expand as the company scaled — instead, every additional shipment was bringing in less contribution than the one before it because compute costs were rising faster than revenue per parcel.
The internal platform team knew something was wrong but did not have the bandwidth to dig in. Every engineer had a kanban board full of feature work and nobody owned the cluster autoscaler, the right-sizing policy, or the analytics workloads that had quietly moved from a nightly batch to a continuously running stream. Half the EKS nodes were running at 18% CPU utilisation; half the EMR clusters were sitting idle between runs that had been on-demand pricing for nine months.
We treated this as a platform engineering engagement first and a FinOps engagement second. The hypothesis was that no amount of dashboarding would fix a system where engineers had no path to act on what they saw. We started by instrumenting cost-per-service in Datadog tied to deployment metadata, then built a self-service rightsizing recommender that engineers could approve with a single click from a Slack message.
On the architectural side, we identified three services — geocoding, ETA prediction, and the parcel-event stream processor — that together accounted for 41% of the compute bill. Each got a focused two-week rebuild: geocoding moved from a brute-force lookup to a tiered cache, ETA prediction moved from synchronous to event-driven with a pre-computed prefix tree, and the stream processor moved from Kafka Streams to a Flink job on Graviton.
The fleet-wide changes did the volume work: Karpenter replaced cluster-autoscaler with workload-aware bin packing; 78% of stateless workloads moved to Spot with PDBs that survived interruption; the analytics tier moved to Graviton-backed Spot fleets with EMR Serverless for spiky workloads. Each engineering team got a weekly cost report tied to their services with a one-click rightsizing approval flow.
“The dashboard didn't fix it. The Slack bot fixed it. Once an engineer could approve a rightsize without leaving a thread, the whole org's habit changed in six weeks.”
Monthly compute spend landed at $1.55M by the end of month five — a 62% reduction against the baseline. The cost curve flattened against revenue, and the CFO presented a revised contribution-margin trajectory at the next board meeting. The three rebuilt services each came in under their target latency budget while running on a fraction of the previous capacity.
Equally important, the FinOps practice stuck. Six months after the engagement closed, the cost-per-shipment metric had continued to fall as engineering teams kept acting on weekly cost reports. AfriLogistics has since hired its first dedicated FinOps engineer to own the platform we built.
FinOps engagements fail when they remain dashboards. They succeed when there is a low-friction path from insight to action — and that path almost always lives in the tool the engineers already use all day. The Slack-bot loop was a deliberately small piece of work that did more for the cost curve than any of the architectural rebuilds. The architectural work was necessary, but it was the cultural rewiring that compounded after we left.
Tech Stack
Services Engaged
Get the long-form case study including architecture diagrams, operating cadence, and the unabridged interviews.
Tell us a little about yourself and we'll send you the PDF.
Let's talk. We'll walk you through what an engagement looks like end to end.
Talk to our team