Technical feasibility prototype
Testing machine learning for trade mark examination
A six-week prototype helped IP Australia encounter what AI-assisted examination might make possible—and where the technology was not yet ready.
Read the case study
Case studies
What changed after people touched the possible future?
The most consequential result of exploratory work is not always a production launch or a headline metric. Sometimes it is a clearer decision, a new organisational capability or evidence that the time is not yet right.
These case studies document the original question, why the answer was uncertain, what Flink built and what the work made possible. They distinguish between prototypes, operational systems and outcomes that can be substantiated.
Selected case studies
Three ways working evidence changed the question.
Together they show three useful forms of progress: deciding not yet, building a capability that endures, and revealing patterns that people could act on.
Long-term data capability
Turning patient experience into visible priorities
A platform operating since 2014 replaced slow annual reporting with recurring, explorable evidence for Victoria's health system.
Read the case studyExploratory analytics prototype
Finding behavioural patterns in supermarket movement
Bluetooth traces, basket summaries and probabilistic modelling became an interactive way to see how store layouts and promotions shaped behaviour.
Read the case studyIP Australia · Technical feasibility prototype
Testing whether machine learning could assist trade mark examination.
IP Australia wanted to understand whether emerging AI techniques could reduce the burden of first-pass investigation without removing examiner judgement. A six-week engagement made both the opportunity and the technical limits concrete.
Six-week engagement
End-to-end feasibility prototype
Text, image and classification analysis
The question
Trade mark examiners were manually comparing new applications with existing text, images and classifications. Basic keyword search helped, but the archive included scanned documents and inconsistent structured data. With application volumes growing, IP Australia wanted to know whether machine learning could surface likely similarities and classification issues early enough to focus expert attention.
The aim was assistance, not autonomous examination. Patent applications were deliberately left outside the first scope so the technical question could be tested honestly.
What Flink built
Flink created the supporting pipeline as well as the prototype: extracting and cleaning content from scanned PDFs, turning it into structured records, mapping trade marks to classifications, and training text and image models on the available material.
The resulting end-to-end tool accepted the text and imagery from a new application, surfaced potentially related marks and provided signals about the proposed classification. Much of the work involved building custom utilities because an integrated off-the-shelf path did not yet exist.
What emerged
IP Australia could test the prototype with examples it already understood. The approach showed credible potential, but the data and technology were not mature enough for a dependable operational system at that point.
The prototype was not adopted as a production service. It did the job it was built to do: reveal what AI-assisted examination could become and provide evidence that a larger investment was premature. The answer was a qualified not yet, rather than a manufactured success story.
Department of Health Victoria and Ipsos · Long-term data capability
Turning patient experience into visible priorities.
Since 2014, the Victorian Healthcare Experience Survey platform has turned patient responses into recurring, explorable evidence for Victoria's health system.
Operating since 2014
From annual PDFs to recurring results
Designed, built, hosted and evolved by Flink
The question
Patient feedback had been treated largely as an annual satisfaction-reporting exercise. Survey data was analysed in spreadsheets and delivered through long PDF reports, often six to twelve months after the patient experience itself.
The Department wanted to shift from measuring satisfaction to understanding experience, shorten the feedback loop and give health services a practical way to explore their own results. The answer had to accommodate lengthy surveys, early paper-and-OCR workflows, varied levels of data literacy and a growing longitudinal dataset.
What Flink built
Working with Ipsos, Flink designed, built, hosted and has continuously evolved the complete reporting platform. It ingested new survey data on a recurring basis and replaced static reporting with custom interactive visualisations that allowed services to move from statewide patterns into the cohorts and questions that mattered locally.
The analytical layer grew with the programme: machine-learning classification of free-text comments, cohort and key-driver analysis, trend modelling, and human-readable insights generated from statistical results. The interaction design evolved alongside the methods so that sophisticated analysis remained usable by non-specialists.
What emerged
Results that once appeared in an annual document became available in recurring reporting cycles—initially quarterly, with some reporting now monthly. The original turnaround fell to less than three months, including the time needed to contact patients, collect responses and process each reporting period.
The platform has operated for more than a decade and its lineage has been adapted for other patient-experience programmes and jurisdictions. Its substantiated contribution is a durable, faster and more analytically capable feedback system. It would go beyond the evidence to claim that the platform alone caused improvements in patient care.
Consumer shopping experience · Exploratory analytics prototype
Finding behavioural patterns in supermarket movement.
An Australian Bluetooth-tracking company had thousands of location coordinates but no way to see what the movement meant. Flink built the visual and probabilistic tools needed to turn those traces into testable retail evidence.
Interactive movement visualisation
Probabilistic journey modelling
Store-specific behavioural evidence
The question
Bluetooth beacons attached to shopping trolleys produced a sequence of x-y coordinates as people moved through a New Zealand supermarket. The company could estimate visit length, but wanted to understand dwell, traffic, route patterns and the relationship between movement, promotions and basket value.
The data did not contain an answer by itself. It had to be reconstructed as journeys, related to the physical store and connected carefully with summarised basket and sales data.
What Flink built
Flink created a processing pipeline and an interactive web tool that animated movement traces over the store map. Heat maps and time-based views made patterns visible, while Markov and hidden Markov models represented the likely sequences within thousands of individual journeys.
Cohort analysis connected dwell and movement with summarised basket value, product categories and promotion periods. This allowed the team to ask not only what a group bought, but where people with a particular basket profile travelled and lingered elsewhere in the store.
What emerged
The prototype revealed store-specific patterns around aisle ends, high-traffic areas, dwell and the routes taken by higher-spending cohorts. It gave the supermarket evidence for tuning product placement and promotion locations beyond general retail folklore.
The tracking company used the work as it expanded into further retail settings and later applied related proximity analysis in other domains. The value was not a universal rule for every store. It was a way to discover which patterns were present in this one—and which were worth acting on.
A strong case study follows what the prototype made possible—even when the possibility changed shape.
Flink publishes only evidence it can substantiate: what people experienced, what became clearer and how the work changed the direction that followed.
Discuss an AI opportunity