Four projects, told the way they actually went. Longer write-ups of some of them live in writing.
The Met Office compresses eight separate hazard models into one public matrix with two axes, likelihood and impact. The compression is necessary, since nobody wants to parse eight risk models before deciding about an umbrella, but it costs personal relevance. Everyone glancing at the app gets the same signal whether or not it applies to them, and yellow warnings, which are the majority of all warnings issued, are the ones people ignore most consistently.
We used COM-B to locate the breakdown and found two capability problems: limited attention under high cognitive load, and difficulty interpreting risk and relevance. High alert frequency paired with low personal relevance produces habituation, which the literature calls warning fatigue and the cry wolf effect.
Our intervention was personalisation as a filter. An onboarding survey about commuting, flood exposure, and time spent outdoors routes users into back-end packs, and each pack filters yellow warnings against thresholds the user set themselves. Red and amber stay untouched, which also limits the false-security risk we mapped on the ripple diagram. Self-determination theory is the backbone, since a filter someone shaped themselves gets processed deliberately rather than screened out automatically, and we checked the design against EAST.
The brief asked us to find a small business in London with a real problem and design an AI intervention for it. We chose Regency Cafe deliberately, because a vintage legacy brand is the hardest possible case for that brief. Their whole selling point is a refusal to modernise, so any recommendation had to survive the objection that it would erode the thing customers actually come for.
I led the client conversations. The owner was sceptical, and his scepticism was well founded rather than a barrier to be overcome, so most of the work turned out to be finding a framing in which automation protected the brand instead of diluting it. That reframing, more than the technical design, is what made the recommendation something he could imagine saying yes to.
For an edtech client I worked on brand and market positioning, running audience segmentation and a competitive landscape analysis to support a leadership decision. For a social mobility charity I worked on widening access to work experience, which meant community needs assessments and key informant interviews to find the structural barriers rather than the stated ones.
Both engagements taught the same thing from opposite directions. The analysis is the easy half. The hard half is landing on a recommendation the client can carry out with the people and the budget they actually have.
I analysed performance across more than fifty digital campaigns for effectiveness and share of voice, and turned the output into decks for stakeholders across every line of business and several EMEA offices. I also led the transition to a new campaign booking system, which took more than two hours of manual processing out of the team's week.
The useful lesson was about what happens to research after it leaves the researcher. The findings that changed the next campaign were the ones tied to a decision somebody already had to make by a date. The rest were read and filed politely.