The reason we take on work in Sha Tin is that the businesses here tend to be sharper about what they want than the brief lets on. AI fraud detection for a Sha Tin team almost always ends up looking different to AI fraud detection for a downtown Auckland one.
AI fraud detection that lives in your stack, not on a vendor's roadmap. Shipped from New Territories.
Our goal is to give Sha Tin businesses a three-day weekend, so people can spend more time with their families and the people they love :)
What AI fraud detection actually does
Pattern-watching AI for refund abuse, chargebacks, fake reviews, employee fiddles, and odd supplier invoices. Flags weirdness early - before it's a real problem.
- 01 Learns your normal patterns and flags outliers
- 02 Daily anomaly report, not a constant alert flood
- 03 Explainable scoring so you can act with confidence
- 04 Integrates with Xero, Shopify, and POS systems
Built on: Claude DuckDB Postgres Vercel
Without Kiwi Dynamics
With Kiwi Dynamics
What you actually get
Every engagement is scoped and quoted up front. This is what is in the box.
- A working pilot, in productionNot a prototype on someone's laptop. The first slice of AI fraud detection runs against real work within weeks.
- Your data stays yoursIt runs on your accounts and your tools. If we part ways you keep the system and everything in it.
- The workflow mapped before codeWe write down what good looks like for Sha Tin businesses first, so nobody is guessing at handover.
- Support after it landsThe people who built it stay reachable when the business changes shape around it.
How AI fraud detection compares
The two things most businesses do instead, and where each one runs out.
| Hiring for it | An off-the-shelf tool | Kiwi Dynamics | |
|---|---|---|---|
| Fit to how you work | Fits perfectly, costs a salary | You bend your process to suit the tool | Built around the workflow you already run |
| Time to something useful | Immediate, and permanent | Quick to switch on, slow to make fit | A working slice in weeks, then hardened |
| Who owns the data | You do | The vendor, on the vendor's terms | You do, in your own accounts |
| When it breaks | That person sorts it, if they are in | A support queue and a ticket number | The people who built it |
| What it costs | A salary, every year, forever | Per seat, forever, used or not | Scoped and quoted up front |
Our field notes from Sha Tin builds.
- Sha Tin anchors the New Territories' manufacturing, logistics and science park economy - AI here means industrial-scale operations work, not office admin.
- Hong Kong Science Park and a broad manufacturing and logistics base serving cross-border trade with mainland China sit here. Businesses want AI that handles scheduling, compliance and cross-border logistics without becoming the bottleneck.
We work with teams across Sha Tin: Science Park · Tai Wai · Fo Tan · Ma On Shan.
Talk to us about this →
How we build AI fraud detection for a Sha Tin team.
We scope narrow, ship a working pilot, then harden it into production. The first slice is the highest-leverage workflow for your Sha Tin business, so value lands before the build is finished. AI fraud + anomaly.
Talk to usHow the work runs
- Find the one workflowWe look for the job costing Sha Tin businesses the most hours or the most leads, and deliberately ignore the rest for now.
- Ship a working sliceA narrow version goes into production in weeks, against real work, so the value shows up before the build is finished.
- Prove it, then widenWe measure it against what the work cost before. If it does not pay for itself, we say so rather than scaling it.
- Harden and hand overLogging, fallbacks and a real handover, so it keeps running when we are not in the room.
The outcome for Sha Tin teams
If we build the right slice first, Sha Tin teams feel the difference inside the first month. Recovers 3-5x its cost in caught fraud within 6 months.
Not your typical AI agency.
Honest about what AI can and cannot do
Ships the one workflow that pays for itself
Hours given back, never the size of the invoice
*Every engagement is scoped and quoted up front. Results vary by workflow and business.
FAQ
How fast could we have AI fraud detection in production?
Eight to ten weeks for most Sha Tin businesses. Faster if your data is in good shape and slower if we're untangling a legacy integration first. We'll give you a realistic number on the scoping call rather than the optimistic one.
What's the smallest engagement you'd take on?
A two-week paid discovery for Sha Tin businesses that aren't sure whether the build is worth doing at all. You get a one-page write-up of what we'd build, what we'd skip, and what it would cost. About 30% of those discoveries end with us recommending you don't proceed.
Anyone else in this space using AI fraud detection?
Plenty. Recovers 3-5x its cost in caught fraud within 6 months. The interesting question is rarely "does it work" - it's "is your team ready to use the output." That's what we'd scope on the call.
What if our Sha Tin doesn't have any data ready?
Most don't. Getting the data into shape - ingestion, cleaning, the lightweight contracts you need before any model is useful - is part of the engagement. For AI fraud detection specifically, we typically run that work on Claude, DuckDB, Postgres, Vercel and assume messy starting conditions from day one.
One short call.
Tell us what you're trying to fix. We'll come back inside a working day.
Talk to us about this
Tell us what you're trying to do and we'll reply with how we'd build it - no obligation.