Keeping research discussion and AI answers attached to the data

Getting oriented

The page. Every dataset in Fuze has a page of its own. It says what the data is and who owns it, holds discussion about the dataset, and contains the research table.

Why it matters. The table is one part of that page, so a conversation has two places to attach: the dataset as a whole, or a single value in the table.

Research context. Informational interviews with scientists helped me understand the broader research workflow and what context mattered when they interpreted data or evaluated an AI response. I used that context in the prototypes I later tested.

Discussion that stays on the data

Shipped

The problem. Researchers questioned values and agreed on meaning in email, chat, and spreadsheet comments, away from the data. When someone wrote "this number looks off," nobody could be sure which number.

The solution. Annotations put that conversation on the table:

• Turn annotations on, and blue indicators mark every cell that has a thread.
• Select a cell and its thread opens in a sidebar, so the table stays in view while people read and reply.

1. 1Annotations On Blue indicators mark every cell that has a thread.
2. 2Select a cell The cell is outlined and its annotations open in the sidebar, beside the table.
3. 3Expand replies Each reply shows the author's role; the dataset owner is marked Owner.
4. 4Reply The composer takes @mentions: “Type anything, or @Mention anyone.”
5. 5Close The table and its indicators stay in place.

Replay View still

Scroll sideways to see the whole page.

Brian facilitated one moderated usability study in July 2024 using a detailed script with four internal scientist participants across four Figma prototype scenarios.

Designed from a moderated study I facilitated. In July 2024, I used a detailed script with four internal scientist participants across four Figma prototype scenarios. Participants wanted to know who was speaking and to pull colleagues in. The final designs added two things in response:

• Each author's role on every reply, and an Owner badge for the dataset owner.
• @mentions in the composer.

Two kinds of discussion, named apart

What we tested. The Figma prototype had two discussion surfaces, one for the page and one for a table cell, and both were called Annotations.

What we saw. In the moderated study, some participants couldn't tell which of the two they were using.

What changed. Each surface got its own name, matched to where it lives:

• Discussion about the dataset became Comments, in an accordion on the page.
• Notes on a specific value stayed Annotations, in the table's sidebar.

Scroll sideways to see the whole sheet Full view

The system underneath. One reusable thread system carries both surfaces. The same role, Owner, and reply treatments work in the wider page Comments area and the narrower cell Annotation sidebar.

Scroll sideways to see the whole sheet Full view

---

Ask AI: designing around uneven model output

Shipped

The technical context. Ask AI, labeled AI Insights in the product, used a RAG-based retrieval approach backed by Llama. The ML and search team owned retrieval, model behavior, and the technical implementation. My work was the interaction layer: how people started, what context they gave the system, what happened while it worked, and how they checked the answer.

The constraint. Open-ended prompting exposed uneven model and retrieval performance. A blank chat box put too much burden on the user: people had to know what to ask, understand what context was active, and decide whether an answer was grounded enough to use.

That pushed the UX toward a more guided, inspectable interaction:

• Guide the start: Suggested prompts gave people useful entry points instead of an empty box, with contextual actions around the message bar.
• Make context explicit: Before analysis, people chose which sources Ask AI could read.
• Keep waiting legible: A progress state showed that work was happening while the layout held still and the input kept focus.
• Return evidence with the answer: The response arrived with the documents, datasets, and tables it drew from, including file details such as size and update date.
• Recover from weak context: When the system did not have enough context, the UI suggested a more specific question instead of pretending the answer was stronger than it was.

1. 1Open AI Insights From the dataset page's sidebar, next to the data.
2. 2Choose sources Pick what it can read, then Continue.
3. 3Start from a suggestion Starter prompts instead of a blank box.
4. 4Wait A typing indicator. The layout holds and the input keeps focus.
5. 5Answer and sources The whole answer, then the documents, datasets and tables it drew from.

Replay View still

Scroll sideways to see the whole page.

A text-first assistant. At this stage, Ask AI was primarily summarizing and retrieving text/data rather than producing rich new visualizations. That made the surrounding chat ergonomics more important: guided starts, context selection, clear response states, source visibility, and useful next actions had to carry more of the product experience.

Beyond the core loop. I also explored surrounding message-bar, upload/context, and follow-up action patterns. The public case only treats controls as shipped when the project evidence supports them; familiar chatbot actions are not added simply because they are conventional.

In the file manager. Ask AI also needed a place in the file manager. I laid out three options: an action on each file row, a single entry at the top of the list, or a prompt bar. The available evidence does not establish which option went forward, so the case does not claim one.

---

FuzeAI: a chat-first direction

Prototype Did Not Ship

The ask. Leadership asked me to show where a chat-first Fuze could go.

The concept. Searching, browsing, and organizing all happen through one prompt:

• Threads sit on the left.
• Uploads and files sit on the right.
• Tables appear inside the conversation instead of on separate pages.
• Anything useful can be saved to a collection.

How it was used. It supported alignment and feasibility discussions. It was not built.

---

Where it landed

• Annotations and page comments: Shipped. Discussion moved into the catalog, tied directly to the data it was about.
• Ask AI: Shipped. The interaction layer guided people toward useful questions, made source scope explicit, and returned source-visible answers inside the research workflow.
• FuzeAI: Prototype Did Not Ship. It gave leadership a concrete direction to evaluate instead of a description.

• Competitive pattern research. Prompt starters, source attribution, upload-to-prompt, message-bar behavior, and chat mechanics across adjacent AI tools. Third-party screenshots are not reproduced.
• Informational interviews. Scientist interviews used to understand workflow context and what information mattered when evaluating research data and AI responses. These are context-setting research, not additional usability rounds.
• Upload flow. Empty to complete states, plus a drag-and-drop enhancement.
• Comment flow states. Accordion threads in empty and filled states.
• Notification options. Post-MVP options that did not ship: in-app vs email, snooze, and unread.
• Test prototype. Four wired scenarios, with hotspot hints off to observe natural discovery.
• Research method. Moderated usability study facilitated by Brian · July 2024 · four internal scientist participants · 45-minute remote sessions · four scenarios on a Figma prototype.