Industry: Startups (leisure, travel and tourism), energy (oil and gas), insurance
This is not a single project but several done with considerable interaction with AI. What amazes me is how quickly we've come in such a short space of time.
"Have you designed AI interfaces?"
Yes, I most certainly have, for a large multinational. Although a design system was mandated, it only followed other systems but, based on findings from research and testing, it was adequate but not optimal for our users' needs. Conversations were likely to be long so locating earlier information was difficult; useful items from several threads were hard to combine; repeating work from other threads but with new data might be tricky; sharing information in threads was not so easy.
The familiar problem from when I studied search engines in the 1990s was still here. It was easy to get information but hard to form information into a useful structure was hard. My job was to not just present a standard chat interface. That's easy. Instead, I looked forward to tackle the challenges users were already beginning to encounter.
Due to an NDA, I cannot show this work but suffice to say that (for example) I designed a way to help users locate information before ChatGPT did by a few days.
I've worked on complex B2B SaaS products for advanced professional users and B2C business channels selling a service.
At one company, we used Figma for designs but as good as it was, it could not handle some complex interactions. The interface had to be in code which I can do but Claude is much quicker. Because prototypes were used for usability testing and demonstration, they did not need to be hardened to production quality. With Claude and later Codex, I could design more interfaces and do so more quickly than ever. Prototypes were fully interactive.
Below is a video showing an anonymised interface early prototype designed with Claude Code. This interface is for a complex platform where, as part of a workflow, engineers will select stations for an environmental analysis. Users can search by name, sort and filter stations for inclusion or exclusion, or can apply complex rules. This was used for usability testing and demonstration of concept to stakeholders.
I also used Codex to very quickly design a small leisure website for a client. This was a proof-of-concept and is still ongoing. What Claude helped me to do is refine the design questions down, create the materials according to my instructions and get it up and running, including a language translator in both English and Thai.
I like data analysis but I dislike the interfaces to analytics software. For some small clients, I collected analytics data from GA4, Yandex Metrika, Google Search Console, Yandex Webmaster Tools, Bing Webmaster Tools and some self-developed tools for site analysis. Making sense of the lot was hard and specifying data and dates to each of these in turn was slow.
But how about if I could ask an LLM a question about the business, like "Should Google Business Profile point to the Chinese or English language site?" and the LLM gets the data for me, interprets it, and then I work with the LLM to validate the findings?
So I set up a server that took a script from the LLM using a schema, retrieved the data back to the LLM which then interpreted the findings and worked with me to generate really useful data-driven recommendations based on multiple sources of evidence?
The below screenshot shows it in action. Yeah, it's just a chat window but I got the answer to the above question sooner than it took GA4 to load. Analysing the data was like collaborating with a knowledgeable friend who did all the heavy lifting while I applied my business expertise.
LLMs have been fantastic for desktop research (but always check the sources!) and they help when designing studies (designs, screeners, discussion guides) and can help with analysis. I prefer to write myself because LLMs, well, they're not quite Shakespeare. Neither am I but I aim to be more engaging.