
On a sale day, one slow step in the warehouse system holds up every order behind it.
On past sale days, a Malaysian logistics and fulfilment company's warehouse system kept stalling at one step: wave allocation, where orders are grouped into batches (waves) and matched to stock. Until it finished, pickers could not start.
By the 11.11 campaign in November 2026, that slowness was gone. This is the AI engineering approach that got it there, and how it applies to any slow or older system.
Our AI engineering approach
We do not point an AI at a codebase and hope. Our approach puts three things together, and starts from one rule: measure the live system first, so every gain is proven the same way.
Graph engineering: a map of the system
Graph engineering means building a map of the system. The stored procedures (saved sets of instructions inside the database), the tables and the application code are points on the map, and the lines show what uses what.
With that map, the agent and the engineer can both ask one plain question: "What breaks if we change this?" The answer comes straight from the map, not from whoever happens to remember. That matters most in an older system, where the people who built it may have moved on.
Automation agents: a wider search
Automation agents scan the code and the database statistics at the same time, far more widely than a manual review. They propose several fixes for each problem, and every fix is tested side by side against the baseline, on real data, before anyone chooses.
Engineers in the loop
Our engineers set the goal, review every finding and decide every release, together with the client's IT team. A fix only counts when it gives exactly the same business result as before, only faster. More on that in who reviews AI-generated code.
The loop: how the agents troubleshoot themselves
The part we are proudest of is not any single fix. It is the loop the agents run, round after round, until the slow step is gone.
- Map. The code knowledge graph shows how the whole system connects.
- Detect. Agents read the live database statistics and flag the slowest paths.
- Diagnose. The agent walks the graph to the root cause, and sees what else a change would affect.
- Fix. It proposes several candidate fixes, and an engineer reviews them.
- Test. Each candidate is tested on real data against the baseline.
- Learn. Every result goes back into the loop as evidence. Proven wins go live in phases, and live timings feed the next round.
Here is the loop at work. For the most costly step, the first idea was a large rewrite. Test showed it was slower, so the agent went back to Diagnose with that evidence, and a much smaller change won. And because a sale day can only be measured on the live system, Learn feeds live timings from every release into the next round. Proven fixes went live in phases, lowest risk first, together with the client's IT team, and by the 11.11 peak the slow step was no longer slow.
For any business running an older system, this changes two things. The system gets diagnosed the same way every time, not by whoever remembers it. And every fix is proven against the baseline before it goes near a peak day.
Before vs now: two ways to fix a slow system
Traditional figures are our estimate, not a measured project.
Results at a glance
About 20×
faster average allocation time across the wave-allocation releases since June (our before/after measurement on the live system)
Nearly doubled
allocation speed for the larger online-order (B2C) waves after the September release
11.11
the wave-allocation slowness of previous sale days, resolved by the latest live release
The 20× is the average across every wave-allocation release since June. The September release was one milestone on the way.
What this means for your business
This approach is not tied to warehouses. Any system that slows down at the worst moment, whether an order platform, a billing run or an older in-house system nobody fully remembers, can be mapped, measured and fixed the same way.
Before your next peak, ask your IT team one question: do we have a record of how the live system behaves on a normal day? If not, start there. Then insist on the same three things: a map of what connects to what, a wide search with every finding checked by a person, and every release measured on the live system.
This is how our software development team works, using the methods behind our AI solutions. For another example of measuring first, read how we cut a Malaysian publisher's cloud costs with no downtime.
Frequently asked questions
Can AI find performance problems in an old system?
Yes. Automation agents can search the code and the database statistics far more widely than a manual review, and a code map shows how every part connects. Our engineers confirm each finding before anything changes, so the search is wide and the decisions stay human.
What is graph engineering?
It means building a map of how a system's parts connect: the stored procedures inside the database, the tables and the application code, and which of them call or read the others. Engineers and AI agents use the map to ask "what breaks if we change this?" before touching anything.
What is a self-troubleshooting loop?
It is a repeating cycle the AI agents run: map how the system connects, detect the slowest paths, diagnose the root cause by walking the code graph, propose fixes for an engineer to review, test each one against the baseline, and learn from every result. Only proven fixes go live, and live timings feed the next round.
Do you still need engineers?
Yes. Our engineers set the goal, review every finding, choose between the fix options and decide every release together with the client's IT team. The agents widen the search; people make the calls.
How do you test before a sale-day peak?
Each fix is tested on a full copy of the real data, with many users at once, and must give the same business result, only faster. Fixes then go live in phases, and each step is measured on the live system before the next campaign.
Free consultation
Is your system slowing down on your busiest days?
Tell us what happens on your peak days: where it slows, and what your team sees. We will set up an online meeting to talk it through and share what we would look at first.

