Bespoke Builds
Custom enterprise AI: computer vision computer vision
For enterprise requirements that off-the-shelf can't meet.
Trusted by enterprise teams in NZ and beyond
-
Real-time vision
A production line where defects need to be caught in real time, and no off-the-shelf vision system covers what you actually need to detect.
-
Predictive signal
A fleet of assets whose failure patterns sit in your telemetry data, where you need a forecasting model that gives usable signal early enough to act on.
-
Document extraction
A pile of documents in formats no commercial OCR handles cleanly, where the value is in the structured data you can extract from them.
-
Custom classification
A classification problem at scale where the taxonomy that matters to your business doesn't map to anyone else's.
-
Novel methods
An operational or research challenge where the AI approach itself is part of the work, and you need a partner who picks the right method instead of the default one.
- Something else?
If your problem doesn't fit any of these, it's still worth a conversation. The harder ones are the ones we like.
Four kinds of bespoke work
Computer vision, document intelligence, forecasting, classification. The four kinds of bespoke work we build most, each with proof in production.
-
Computer Vision
Inspection systems, monitoring, and object detection in environments where commercial vision platforms don't fit. We've built these for industrial production lines, security-sensitive operational environments, and edge contexts where data can't leave the site.
A national primary industry body engaged us to build a vision system for defect detection at processing-line scale. Custom hardware design with an industrial partner, ML model development for a problem where off-the-shelf vision didn't apply, and a multi-year rollout replacing manual inspection across the sector.
-
Document Intelligence
Extracting structured data from unstructured documents, packaging, forms and imagery, where commercial OCR or document AI doesn't handle the variability. Often involves multimodal LLM orchestration and custom benchmarking against human performance.
A national data services organisation asked us to automate extraction of structured product data from packaging imagery, work previously done manually by an offshore team. We evaluated multiple multimodal LLM providers, built a custom benchmarking method comparing AI to human accuracy, and the AI is now meeting or exceeding human performance across multiple data streams.
-
Forecasting and Predictive ML
Time-series forecasting, failure prediction, asset health monitoring at fleet scale. Often involves edge ML, signal processing, and integration into operational systems that need usable signal early enough to act on.
A B2B industrial SaaS platform engaged us to build predictive ML for asset failure and lifespan estimation, currently scaling to a target customer monitoring more than 100,000 assets across multiple international sites.
-
Classification and Recommendation
Categorisation at scale, recommendation systems, and personalisation where the taxonomy or behaviour model that matters to your business doesn't fit a generic approach. Often combined with computer vision or document intelligence depending on the data input.
AI-driven product attribute labelling for a major NZ retailer, enabling new browsing experiences (such as “shop by age” for toys) without manual classification at scale.
How we approach a bespoke build
Every bespoke build is different, but the way we deliver them isn't.
-
Start in context
We do discovery in the environment the work will live in. On-site if it's industrial, alongside your team if it's operational.
-
Pick the right approach, not the default
Modern AI offers a lot of methods. We benchmark options against your data and pick the one that fits, not the one we're most familiar with.
-
Build against benchmarks you can see
We define success in numbers at the start and measure as we go. You see the work as it lands, not at the end.
-
Design for the operational reality
Legacy systems, security constraints, human workflows. Integration is part of the build from day one, not a final-stage afterthought.
-
Hand it over so it keeps working
Capability transfer is built into the engagement, not bolted on at the end. Your team can maintain, improve and extend what we hand over.
Examples
Bespoke Builds running in production, across multiple sectors and capabilities.
-
Computer vision
A national primary industry body
AI-driven defect detection on a processing line, replacing manual inspection across the sector.
-
Document intelligence
A national data services organisation
Multimodal LLM extraction of structured data from product packaging, meeting or exceeding human accuracy.
-
Forecasting · Predictive ML
A B2B industrial SaaS platform
Failure prediction and lifespan estimation for asset fleets, scaling to a customer monitoring 100,000+ assets.
-
Edge computer vision
A global industrial services provider
On-premise edge CV for crowd and traffic monitoring at large public events, and foreign-object detection on industrial conveyors.
-
Computer vision · Classification
The Warehouse Group
AI product attribute labelling enabling new browsing experiences without manual classification at scale.
-
Recommendation
The Warehouse Group
AI gift recommendation engine that lifted Christmas conversion.
-
Document intelligence
vWork
AI-driven document extraction for field service operations.
-
Forecasting · Data science · IoT
Autogrow Systems
Development partner for an IoT and data science platform that increases plant yield.