Self-initiated concept project · Not yet shipped · Published on Figma Community
01
WHERE IT STARTED
As a Founder Product Designer at a Bangalore startup, I was the entire design function — research, synthesis, wireframes, UI, stakeholder presentations. Everything. And I kept hitting the same wall.
60%
of design time disappears between finishing research and starting design
Not in interviews. Not in Figma. In the gap between them.
I'd finish five Zoom calls with users, open Figma, and stare at an empty frame. Not because I didn't have data — because I had too much of it, scattered across Notion docs, color-coded tags, and half-remembered quotes. The synthesis was on me. Every time. Entirely manually.
First instinct — build a repository tool
My initial idea was a better Notion for research. A smarter tagging system. A place to organise insights. I spent two weeks sketching this before realising — the problem isn't organisation. Designers already have tools for that. The problem is the gap between organised data and a confident design decision. I had to throw the first direction out entirely.
The real problem isn't finding insights - it's deciding what to act on
A better filing cabinet doesn't help when the connection between evidence and decision is still manual
The tool needed to do the synthesis work, not just store the raw material
02
DEFINING THE PROBLEM
Before designing anything, I needed to know whether this was personal frustration or a real, shared problem. So I ran a focused research sprint in December 2024 with designers and PMs doing this work every day.
"I spend more time organizing my thoughts than I do having them."
Product Designer · Research Sprint
"Relating each quote to the design idea is completely manual. Every single time."
Product Designer · Research Sprint
"I'm not scared the AI will be wrong. I'm scared it will be confidently wrong in a way I can't detect."
Product Manager · Research Sprint · This reframed everything.
Second pivot — the problem isn't accuracy, it's opacity
I went into research expecting to validate "AI synthesis is inaccurate." What I found was the opposite, people don't fear wrong AI outputs. They fear invisible ones. This completely changed the design problem. It's not about making a more accurate AI. It's about making the AI's reasoning visible enough to be challenged.
HOW MIGHT WE?
01
The Synthesis Bottleneck
Designers aren't stuck collecting data. They're stuck deciding what it means. The last mile between evidence and decision is entirely manual.
02
The Opacity Fear
If designers can't trace an AI output back to a source, they won't act on it. Invisibility kills adoption faster than inaccuracy.
03
The Confidence Gap
There's no signal for "this insight came from 4 interviews" vs "this came from one offhand comment." Weight of evidence is invisible.
03
THE SYSTEM
SETU is built on 6 specialized agents working in sequence. The key design constraint: every agent needed a UI moment — a way to make its invisible work visible to the designer.
The architecture didn't follow the design
The design followed the architecture, every screen exists because of something an agent does.
Vector embeddings + RAG
Cross-project intelligence is feasible through vector embeddings + RAG.
SETU retrieves only the most semantically relevant past insights, not all project data.
Mission Brief
The Mission Brief is the human-readable translation of a RAG query.
A technical constraint turned into a trust-building design feature.
THE TWO FLOWS
REACTIVE
Mission Brief
Designer asks → Agents run → Evidence Cards appear. Answers the question: what happens when you know what to look for.
PROACTIVE
Shadow Researcher Flow
SETU monitors continuously → Pattern detected → Alert fires with pre-generated card. Answers: what happens when you don't know what to look for.
04
DESIGN DECISIONS
DECISION 01
The moment a designer decides to trust, question, or reject an AI output.
The standard pattern for AI output cards is: here's the insight, here's a confidence score, accept or reject. That pattern fails the opacity test. A confidence score with no trace to source data is just a number. It doesn't resolve the fear.
EARLY DIRECTION
Dashboard-style metric card. Data without context. Designer sees the claim but has no path to the evidence behind it. Looks clean. Builds zero trust.
FINAL DESIGN
Claim → Confidence badge → Source depth → Expandable evidence rows with exact quotes, timestamps, and pattern weight indicators. Every claim is traceable.
Pattern dots ● ●● ●●●
one-off, recurring, dominant. Designed once, never explained again
Source timestamp
"Zoom · Nov 14 · 00:04:32 ↗" — clickable, goes to exact moment in recording. Trust through access
Confidence badgeas High/Med/Low
not a percentage. A percentage creates false precision. This is honest about what the AI actually knows
4 states
Default (scan) → Expanded (audit) → Disputed (feedback) → Accepted (spec-linked)
DECISION 02
The instinct every designer has — and why it's wrong.
In every sketch, Dispute was a small link. Tucked away. Politely hidden. The reasoning was familiar: "Edge case. Most users will accept." But the research finding directly contradicted this assumption.
THE DECISION . JAN 2026
EARLY DIRECTION
Accept = primary button. Dispute = small grey link at the bottom. Hierarchically invisible. Functionally present but psychologically buried.
FINAL DESIGN
Three equal-weight buttons: Accept · Refine · Dispute. Dispute gets amber hover state and triggers an inline feedback field. SETU stores the correction and recalibrates.
Dispute is not rejection - it's the feedback loop that makes the system smarter
After dispute: card gets amber left border, confidence badge → "Under Review", inline text area appears
Business case: dispute data = the most valuable training signal. SETU collects. Every correction improves future insight quality
DECISION 03
The harder, rarer design problem. And the one that defines SETU's market position.
Reactive AI tools put the entire burden of knowing the right question on the user. The most dangerous insights are the ones you don't know to look for. SETU's Shadow Researcher monitors connected sources 24/7 and surfaces pre-generated Evidence Cards when patterns cross a confidence threshold.
The design tension - too loud vs too quiet
Alert too often → alert fatigue, users ignore everything. Alert too rarely → misses the insight that matters. The threshold design (3+ sources, increasing trend, High/Medium confidence) is the design decision that resolves this tension.
Alert fires only when 3 conditions are true simultaneously: pattern depth (3+ sources) + confidence level (High/Med) + trend direction (↑ increasing)
"Dismiss for 7 days" not "Dismiss forever" - SETU respects attention without losing the signal. Micro-copy as trust design.
The alert panel is not a toast, not a modal - a floating panel with a left violet rail. Calibrated interruption weight.
Pre-generated card is ready before the user sees the alert - zero waiting after tapping "View Insight"
The Autonomy Dial (Observe → Collaborate → Autonomous) lets designers control how much agency SETU exercises. Trust is earned, not assumed
05
WHAT I LEARN
Technical constraints make the best design features
The Mission Brief exists because RAG needs to know which sources to query before running. A constraint became the product's most trust-building moment.
Invisible agents still need design
The Ingestion Agent has no UI. That was a deliberate decision. "Last synced: 2 mins ago" is the entire design for an agent working silently 24/7.
Trust is designed in layers
Every screen before the Evidence Card exists to build trust before the AI speaks. By the time the insight arrives, the designer already knows what the AI looked at and why.
CONCLUSION · SETU · 2026
06
FUTURE ITERATIONS
Figma integration
accepted Evidence Cards automatically update component annotations in the connected Figma file
Confidence calibration over time
SETU tracks accept vs dispute ratio per designer and adjusts thresholds accordingly. The more you use it, the better it knows your standards.
Cross-team intelligence
SETU flags when two designers on the same team are working from contradictory research conclusions