I-Flex / 2026
Figma
Claude
Secondary Research
Product Design
AI Guidelines
Role
Product Designer
Tools
Figma · Claude
Research
2,461 transactions · 45 months
Frameworks
HAX · PAIR · Shape of AI
In short
Problem
Budgeting tools assume a monthly paycheck. For irregular earners, income swung 84× between the best and worst month in the data I studied, so any single-number forecast is wrong before it is read.
My role
Sole designer. Data analysis, flows, UI, prototype and the AI guidelines the product follows.
Timeline
3 weeks
Key decision
Never show a number the app can’t stand behind. Predictions are ranges, user-set numbers are exact, and automation only runs between fences the user sets.
Outcome
A working prototype: Safe-to-Spend, a range forecast with lean-month flags, an income Smoother with user-set floors and ceilings, and an assistant that declines what it can’t answer.
Status
Concept · Prototype · AI
Contents
01 · The problem
Challenge & Approach
84×
Best month vs worst month, across 45 months of real data.
2× a year
Spending outran income, like a season, not an emergency.
1 in 3
Months where a simple forecast missed by 25% or more.
So I‑Flex never pretends to certainty
Forecasts in ranges
Every prediction is a band, never a confident single number.
Automates inside fences
Saving moves only between the floor and ceiling you set.
Shows the receipts
Every claim links to the transactions behind it.
Not replacing judgment. Meeting it halfway.
02 · The evidence
Data Analysis
The swing
between the best month and the worst, in a household that is mostly salaried. For a freelancer, the chart is worse.
transactions studied
months of real data
months spending outran income
forecasts off by 25%+
Swing
Shortfalls
Forecast
Income isn’t a line. It’s a rollercoaster.
Monthly income swung from ₹3,500 to ₹2.9 lakh across 45 months. Any single-number forecast is a lie waiting to be discovered.
Worst month
₹3,500
A lean month
₹18,000
A typical month
₹52,000
A strong month
₹1.4 lakh
Best month
₹2.9 lakh
2,461 transactions from a real household diary, 45 months (Kaggle).
03 · Who it is for
Target Personas
04 · The shift
User Journey
I reorganised Raj’s money month around the app. The journey drops from four phases to three — the famine phase is absorbed into planning — and 7 of the 10 friction actions are automated or answered.
Raj’s remaining work
Confirm, decide and adjust. Never compute.
A crisis phase became a calendar event.
05 · The build
Wireframing
06 · The build
Moodboard & Illustrations
Safe‑to‑Spend: what’s yours to spend, today
The rollercoaster, bridged: income’s shape, and the app spanning its gaps
The buffer, filling: coverage in lean months, not rupees
Lean months, sheltered: protection, not panic
Reference teardown
Three apps solving a neighbouring problem. I took one idea from each, and named what I would not carry over.
Revolut
TOOK
One number owns the screen; everything else is subordinate to it.
LEFT
The dense analytics grid — too much to parse in a ten-second glance.
Oportun
TOOK
Set-aside framed as already handled, so no arithmetic is asked of the user.
LEFT
The fully black-box automation. I-Flex has to show its working.
Cleo
TOOK
Plain-language verdicts in place of charts.
LEFT
The chat-first model — budgeting here has to be glanceable, not conversational.
The system behind the screens
Small enough to hold in your head: one spacing unit, four type roles, eight colours. Every screen in this case study is built from these.
Spacing & padding
4 pt base. Everything is a multiple of it, so nothing is ever eyeballed.
4
hairline gaps, icon to label
8
inside a row
16
card padding
20
screen gutter
24
between cards
Corner radius: 16 on cards, 12 on controls, full round on pills.
Typography
Inter throughout. Four roles — if a new size is needed, the hierarchy is wrong.
₹18,400
Display · 28 / 600
This month
Title · 20 / 600
You can spend this without touching rent.
Body · 15 / 400
SAFE TO SPEND
Label · 12 / 600 · 0.12em
Colour
Green carries the brand, teal is the only accent, and red is reserved — it appears when money is actually at risk.
Ink
#0F1A16
primary text
Green
#184A3E
surfaces that carry the headline number
Teal
#0FC29A
the single accent
Teal soft
#E6FAF4
supporting panels
Grey
#797E7C
secondary text
Line
#E8E8E8
dividers and borders
Amber
#FFEFD3
a nudge, never an alarm
Red
#E0483C
reserved for real risk
How a screen gets built
Good morning, Sid
SAFE TO SPEND
₹18,400
This month
On track
Next month
₹4,200 short
Rent · ₹12,000
SIP · ₹5,000
Home
Plan
Ask
Electricity · ₹1,800 · due 12th
Blocks
Every element gets a box. No sizes, no alignment yet.
07 · The product
Chapter 1 · Forecasts that tell the truth
Lean month ahead?
The app never issues instructions about your life. It prices the trade and lets you decide. “You will need to” is a command; “frees ₹1,040” is an offer.
Comfortable path ahead?
The Forecast page is volatility‑responsive. For a steady earner it quiets down to a three‑second glance: the same components, turned down.
The AI never says “cut this”
It prices the trade (“skipping 1 frees ₹130”) and connects it to a goal you already have.
Calm about good news
The UI changes when the months ahead are comfortable, as calm about good news as it is early about bad.
Out‑of‑scope questions get a decline plus a bridge. The AI never pretends to expertise it lacks, and never leaves you at a dead end.
08 · The product
Chapter 2 · One number answers the daily question
Safe‑to‑Spend
Computed after commitments and buffer, not the bank balance, which lies to irregular earners. Forecast bars ahead are ranges, not promises.
Tap any month, see the working
Past months show their receipts. Future months explain their uncertainty. A forecast that shows its ingredients can be corrected, and trusted.
09 · The product
Chapter 3 · Ask
1
It starts with a contract, not a greeting
The first message states what the AI can see and promises receipts with every answer. No wave emoji, no personality.
2
Chips are promises
Every suggested question maps to an answer the system can render with evidence.
3
Noticing without judging
“Food delivery ran ₹2,340, about double your usual.” Spending is data, never a personality trait.
4
Declines come with a bridge
Asked about crypto, the AI names its boundary and hands over what it does know: how much is genuinely spare.
10 · Craft
Micro-interactions
Each one does a single job: confirm a tap, point at a change, or explain where something came from. These demos are live — press and tap them.
Move ₹6,000 now
Press it
Press confirms the tap
Press the button. It dips to 95%, the shadow tightens, then it settles. One short receipt that the tap landed — nothing bounces.
₹41,000
Jul
₹46,000
Aug
₹28–46k
Sep
₹30–52k
Oct
Press a month
A month names itself
Press and hold any month. It rises, its fill casts a soft shadow, and that month’s figure appears above it — so the chart answers before it opens anything.
SAFE TO SPEND
₹42,600
September may run lean.
Plan ahead →
Press to run a forecast
The lean month asks once
Press the card to run a forecast. The amber line fades in once, says what it found, and then stops asking. Press again to reset it.
Tap a month
September — lean month
Project work swung ₹5,200 – ₹45,000, so the range is wide.
Close
The sheet rises from the chart
Tap a month. The chart recedes and the sheet travels up from the bottom edge, carrying that month’s explanation. Close sends it back the way it came.
Home
Forecast
Plan
Switch tabs
The tab bar carries you across
Switch tabs. The green pill travels to the new tab and the icon colour and label weight change with it, so the movement does the explaining.
SAFE TO SPEND
₹42,600
after commitments & buffer
Apply the plan
Numbers travel, they never jump
Apply the plan. ₹42,600 becomes ₹36,600 in place, and the caption changes with it, so you see a decision take effect rather than a page reload.
11 · The stance
AI Principles
I‑Flex uses AI in exactly three ways
It never chats. It answers when asked, and stays silent otherwise. Built on Microsoft’s HAX guidelines, Google’s People + AI Guidebook and the Shape of AI pattern library.
“An AI that knows when to stay quiet is worth more than one that always has something to say.”
12 · The close
Conclusion
What came out is a small grammar, applied everywhere: predicted numbers are ranges, user-set numbers are exact, and automation runs only between fences the user sets.