Logistics
−18%
Empty-run kilometres
Route utilisation reporting
Fleet telematics joined to order data, surfacing empty return legs to dispatchers in a daily operational view.
Home / Work
The problem, the build, the number that moved.
Client names are withheld under NDA, but the sectors, systems and outcomes are exactly as they happened. We'll walk you through any of these in detail on a call.
0
Senior specialists on every job
0wk
Median time to first value
0
Clouds we deliver on
0%
Code in your repo, always
Case 01 · Manufacturing group
Cutting a six-day month-end close to one
A multi-plant manufacturer running production, quality and sales on three disconnected systems, with monthly reporting assembled by hand in spreadsheets.
The problem
Finance spent the first week of every month exporting from the ERP, pulling quality logs from a plant-floor system, and reconciling both against a sales spreadsheet. By the time leadership saw the numbers, the decisions they informed were already three weeks stale. Nobody fully trusted the output, so plant managers kept their own private trackers in parallel.
What we built
A governed warehouse fed by nightly pipelines from all three systems, with automated reconciliation checks that flag mismatches before anyone opens a report. On top of it, two dashboards: a plant operations view that floor managers open every morning, and a finance close pack that generates itself.
What changed
The close now takes a day. The private trackers disappeared within two months — a better signal of trust than anything we could have measured directly. The same warehouse later supported a scrap-rate forecasting model.
6 → 1DAYS TO CLOSE
~90hSAVED PER MONTH
3 → 1SOURCES OF TRUTH
Case 02 · D2C retail brand
Seeing churn early enough to do something about it
A subscription-adjacent consumer brand with strong acquisition but no visibility into which existing customers were quietly drifting away.
The problem
Retention campaigns were sent to everyone who hadn't ordered in 60 days — by which point most had already switched. Two previous attempts at a churn model had stalled because customer records were duplicated across the store, the helpdesk and the email platform, and no one had reconciled them.
What we built
We spent the first three weeks on identity resolution and an agreed definition of an active customer. Only then did we train a model. It scores the base weekly and writes at-risk accounts back into the CRM with the top contributing reasons, so the retention team knows what to say, not just who to call.
What changed
Repeat purchase rate rose eleven percent over two quarters, and campaign volume dropped because targeting got sharper. The client's own analyst now retrains the model on a schedule we documented.
+11%REPEAT PURCHASE
−34%CAMPAIGN SPEND
5 wksTO PRODUCTION
Case 03 · Financial services
Ending 30 hours a week of manual rekeying
A mid-sized lender whose analysts retyped figures from client PDF statements into spreadsheets before every credit decision.
The problem
Four analysts spent most of their week transcribing bank statements in dozens of layouts. Transcription errors surfaced late, occasionally after a decision had been made, and the team could not scale volume without hiring.
What we built
An extraction pipeline that reads statements, normalises them into a standard schema, and routes anything below a confidence threshold to a human review queue. Crucially, we designed the review step first — an extraction system without a good correction workflow just moves the problem.
What changed
Roughly thirty analyst hours came back each week and were redeployed onto credit assessment. Error rates on extracted fields fell well below the manual baseline, and the audit trail satisfied their compliance review.
30hSAVED WEEKLY
94%AUTO-EXTRACTED
4×VOLUME CAPACITY
Also in production
Shorter engagements.
Healthcare
21 days
Audit-ready reporting
Compliance data warehouse
Consolidated clinical and administrative records into an access-controlled warehouse with a full lineage trail.
B2B SaaS
1 number
Agreed across teams
Metric layer rebuild
Sales, finance and product each reported different ARR. We defined it once, encoded it, and retired eleven conflicting reports.
Want the detail behind any of these?
We'll walk you through the architecture, what went wrong along the way, and what we'd do differently. Then we'll tell you whether your situation looks similar.
Replies within one business day · No obligation