Search Sensei is a knowledge-discovery platform built for organisations that need to locate, understand, and reuse information at scale. It sits at the intersection of search, AI, and agentic capabilities, designed to surface the right answer confidently, not just a list of links.

AI can now answer questions directly, but most interfaces leave users with more questions than answers. Three friction points kept surfacing across existing products:

Search Sensei needed a production-ready search experience that any website could plug in, generic enough to work across industries, sophisticated enough to handle AI responses with confidence.

Our job was to design it end-to-end and hand it off to developers in a format they could build from directly

These were the final deliverables

Search Sensei powered NAB's existing search, but it was purely traditional, no AI involved. That's exactly the gap we were designing to close.

We followed a Design Thinking approach, starting with understanding the space before touching any screens. Every decision from here was grounded in what we observed, not what we assumed.

We started by analysing three products: Google, Bing, and NAB, because they represent different ends of the search spectrum: consumer-scale AI search and a real-world example of enterprise website search. For each, we mapped out every element and interaction in FigJam, screenshotted and annotated key UI patterns, then sorted them into what worked, what didn't, and what was worth adapting.

What we decided to carry into our design:

With our research locked in, we moved into wireframing: working out layout, hierarchy, and screen flow before touching any visuals. We covered the full search journey: the opening state, predictive suggestions, the loading state, search results with AI and traditional views, and the sources and citations detail screen.

We moved into mid-fidelity to start making real design decisions, translating wireframe structure into something you could actually react to, with real content replacing placeholders to get a true feel for the experience

After sharing our initial concept with our client and lecturer, four things came back consistently:

• The AI mode toggle wasn't distinct enough, users couldn't tell it was doing anything different
• Results felt overwhelming, too much text and information competing for attention
• No clear visual hierarchy between AI-generated and traditional results
• Needed NAB branding applied to ground the prototype in a real context

Here's how the final experience comes together, screen by screen.

The default state

The first thing users see when they open search. Trending searches for those who aren't sure what to ask, recent searches and favourites in the sidebar for those who are.

Search results

Results are organised by relevance, not recency. The most useful answer surfaces first, followed by a primary card and supporting results below. Every result is tagged by category so users can orient themselves at a glance.

AI Chatbot

When a search result isn't enough, users can keep going. The chatbot lets users ask follow-up questions and dig deeper, without losing the context of where they started. A “Back to search results” link keeps them oriented the whole time.

The same search query, two completely different experiences. Here's what changes when you choose between:

Traditional modeAI mode

Clean, familiar results. A primary card surfaces the most relevant content at the top, followed by supporting cards and a list of additional results below. No AI involved, just the most relevant content, clearly organised.

Results are ranked by relevance and rendered into four distinct levels, each designed to match how much detail that result deserves.

AI summary card

Only appears when AI mode is on. Generates a synthesised response with a confidence tag, sources, and suggested follow-up questions.

Primary card

The highest-relevance result. Full-width with an image, title, and a CTA button. Always the first thing users see when AI mode is off.

Normal card

The next tier, displayed as a three-column card row with title, description, and category tag.

Lists

Remaining results as compact rows for quick scanning.

Three things I'd carry into every project from here: