Portfolio
AquaDeep: AI Personal Shopper for Sustainable Dive & Surf Gear
STATUS: This is a Master's exercise in hypothesis testing and experiment design it isn't headed for production. What it's meant to show is the discipline behind it: finding the assumption the business actually depends on, designing the cheapest test that gives a real answer, and being honest about what a projection can and can't tell you before the test has run.
The Story
Technical diving gear tends to be ugly, while surf gear is often generic. Yet both attract a demanding, sustainability-conscious buyer who researches extensively before purchasing and still struggles with technical jargon, uncertain fit and thermal thickness, and unverified sustainability claims.
AquaDeep addresses this with an AI Personal Shopper that recommends the right garment based on skill, activity, destination and water conditions, alongside a verification seal for brands with credible sustainability claims.
The Role
Earlier phases covered the user persona, Jobs-to-be-Done, and Business Model Canvas, which provided the foundation for the hypotheses below. My focus this last part was mainly on experiment design, including the Wizard of Oz approach and why it was preferable to a Fake Door test, as well as the AI-simulated projection and its transparency framing.
Key Decisions
Focusing on the riskiest assumption. Earlier research had already validated the customer and problem, so the main uncertainty was the solution. We broke this into three risks: whether users would trust an AI recommendation enough to buy without trying the product, whether it could reduce the sector’s 24.4% return rate, and whether matching manufacturer sizing with real-time ocean conditions was technically reliable.
Choosing a Wizard of Oz test. A Fake Door would only measure clicks, while the key question was whether users would trust a personalised recommendation. We therefore used an existing rule-based prototype to simulate the AI and stopped the funnel after email capture, replacing checkout with a waitlist. This allowed us to test purchase intent at almost no additional cost.
Separating projections from real results. Before running the experiment, we used an AI-generated projection to estimate funnel performance and sanity-check the design. The figures were clearly labelled as projections, not results, and would need to be validated against the real experiment.