SparkAI
/
2026
A ground-up rebuild of an enterprise AI workspace — taking a dark, card-cluttered chat product and turning it into one input-first canvas where choosing a model, speaking a prompt and reading the receipts all happen in the same place.
Who was on this?
Me (Product Designer)
Role
Product Designer
Scope
IA · UI · Interaction
Platform
Web Application
Year
2026
My role and summary
I took an enterprise AI chat tool that had grown in every direction at once and rebuilt it around the one thing every session actually begins with — the input. Model switching, attachments, voice and dispatch all collapsed into a single container anchored to the middle of the canvas. Chat history got a spine. Answers got somewhere to keep their receipts. Everything that wasn’t the exchange between prompt and response got cut.
The measure for every decision was the same: does this shorten the distance between what the User means and what the system gives back?
the input field, walking back to the centre of the canvas after a long absence
Fig. 1 — The redesigned workspace. One input container on the canvas; history grouped into semantic date clusters rather than a flat list.
Challenge, purpose and definition
What’s actually wrong with most AI workspaces?
Enterprise AI tools tend to accumulate. A model gets added, so a card goes on the home screen. A feature ships, so a button joins the tray. Nothing is ever obviously wrong, and then one day the prompt field — the only element every single session touches — is the smallest thing on the page, pinned to the bottom edge, below four cards explaining models the User already knows.
Problem
The one thing every session begins with was the hardest thing on the screen to reach.
everything on display except the one thing you came in to use
I pulled the legacy build apart, and here’s what I found
01
An isolated input zone, detached from the controls that govern it — the model picker lived in the top bar, attachments in a tray, the field itself at the bottom.
02
Static model cards holding the centre of the canvas, describing Auto Pilot, Gemini 1.5 Pro, Llama 3.1 70B and Runway to people who switch models twelve times a day.
03
Chat history with no semantic grouping — nine consecutive threads all reading “1- Greetings & Salutations…”, which is a list that technically works and practically doesn’t.
Fig. 2 — The legacy build. Model cards own the canvas, the input sits at the bottom edge, and the history repeats itself nine times.
my honest review of centre-screen model cards
So I defined what the redesign had to do
People running six models a day need one place to type, one place to look, and no decisions in between.
The design approach
Consolidation first
The model cards came off the canvas entirely. In their place, a single input container sits at the optical centre of the screen and absorbs everything that used to be scattered around it — the attachment control, the model selector, the microphone, the dispatch button. One footprint, four jobs.
Chat history moved into a structured sidebar and picked up chronological grouping — July 2026, June 2026, April 2026, December 2025 — so locating a prior session became a matter of remembering roughly when, not scanning a flat column of near-identical titles. Vector Vault and My Data sit above the fold, where the work actually starts.
one container. no exceptions.
what the model cards said, on the way off the canvas
Model and voice
Both live inside the same footprint
An in-line switcher with search surfaces the whole catalogue — Auto Pilot, Amazon Nova Micro, Azure ChatGPT 3.5 and 4o, ChatGPT 4 Turbo, 4.1 and 4.1 Mini — without the User ever leaving the input context. The default is labelled as the default, so the common path needs no decision at all.
Voice renders as a live waveform inside the field rather than over it. Recording state, elapsed duration and dispatch occupy exactly the footprint the text cursor just vacated. No modal, no overlay, no state to dismiss before you can type again.
Navigation
It collapses when you don’t need it
Navigation is multi-tier and independently collapsible. The primary rail compresses to icons; the history panel folds away on its own. At full collapse the canvas is almost entirely input and answer, which is the point.
Per-conversation actions — Pin Chat, Edit, Delete — moved into contextual hover menus instead of permanent affordances. And a Prompts panel on the right edge lists every prompt in the session in sequence, so a long thread stops being something you scroll and becomes something you navigate.
Fig. 4 — Collapsed icon rail, expanded first-tier navigation, the per-conversation hover menu, and the sequential prompt timeline.
the prompt timeline, keeping score of a very long session
Observability
When an answer needs receipts
A complex AI response is three different things at once, so it segments into three views. Answer carries the structured content, comparative tables included. Sources exposes the reference links, clickable, for anyone who needs to verify before they forward. Details surfaces the execution metadata — runtime, model attribution — so “Executed: 17.40s” is available to the people who care and invisible to the people who don’t.
In an enterprise context this is the difference between an answer and a usable one. Full observability, no visual overload.
Here’s what it looks like
Fig. 5 — Segmented output. Structured answers with comparative tables, and execution metadata one tab away.
Saif Khan — Product Designer
Impact and learning
Impact
Fewer decisions per prompt. Consolidating model, attachment, voice and dispatch into one container removed the cross-screen hunt that used to precede every message.
History you can actually re-enter. Chronological clustering replaced linear scanning, so returning to a session from three months ago is a date recall, not a reading exercise.
Answers that survive scrutiny. Sources and execution details are one tab away, which is what makes an AI output forwardable inside an enterprise.
What did I learn from this?
Subtraction is a design move. The cards weren’t badly designed — they were well-designed things in the wrong place. Removing them was the largest single improvement in the project.
Consolidation beats arrangement. I could have spent the whole redesign moving controls closer together. Putting them in the same container was the version that actually worked.
Enterprise users want the metadata, just not constantly. Runtime and model attribution matter enormously — right up until they’re on screen all the time.
me, when the canvas finally had nothing left to remove
Next case study
Techxpo 2.0