AI-first product designer
I'm Rohit, an independent product designer for ambitious, early-stage B2B and B2C teams building something new. I've designed products for YC-backed teams across fintech, compliance, data analytics and consumer apps.
How I work
Design tools bring the craft, AI brings range and speed, and every decision stays human.
Before any screens, I study how the work happens today: interviewing the people who use the product, reviewing support tickets and analytics, and using AI to synthesize what I find and surface patterns that are easy to miss.
Working in design tools and AI together, I turn ideas into clickable flows quickly, so the team reacts to something concrete instead of a static screen or a description.
I observe where users hesitate and which workarounds they create, and use AI to analyze sessions and feedback at scale. The judgment about what to change stays with me.
Every edge case is resolved and documented, with tokens and components engineers can build from directly, without needing clarification.
Selected work
Revalgo
Sales Order Automation
MercuraYC 25
AI Procurement
Mira Wealth
Wealth management platform
Your project
Currently accepting clients
AI changes how much we can explore, not who is accountable for the result. I use it to test more ideas and cover more ground, and I stay with the project through the difficult middle to make sure what ships is right.
Engage
I work with two clients at a time, three at most, so every team gets consistent weekly attention.
One hard problem, taken from an unclear idea to a clickable, testable flow your team can react to.
Design and development running together week after week, with design tokens and components rather than a folder of static files.
From early idea to a design your engineers can build from. Quoted upfront, so if the work takes longer, that risk is mine.
Still got questions?
I design products and the brand identity around them, for B2B and B2C teams. I add the most value when the direction of the product is still being defined.
AI is part of my process, not an add-on. I use it alongside design tools for research synthesis, exploring directions, prototyping and documentation, and I review everything myself. The result is broader exploration and faster iteration, with accountability for the final design staying with me.
I work with a maximum of three clients at once, and usually two. That limit is what allows me to keep a consistent weekly cadence with each team.
I work with teams in US and European time zones and schedule calls to overlap with your working day.
Yes. Design and development run in parallel, so what ships matches what was designed. You receive design tokens and production-ready components rather than a folder of static files.
Ongoing partnerships can be paused or ended with 30 days' notice. Fixed-scope projects are quoted upfront, and the price does not change if the work takes longer than estimated. That risk is mine.
Yes, before any design begins. I speak with your users and review whatever analytics already exist. These conversations often show that the underlying problem differs from the one I was originally brought in to solve.
Sprints are paid upfront. Fixed-scope projects are billed 40 / 30 / 30 across the project milestones, and ongoing partnerships are billed monthly in advance. Payment is by bank transfer, to either a US or an Indian account, whichever is more convenient for you.
Book a call and let's talk.
Book a callOr email → rohitsurvase.creates@gmail.com
Get in touch
A short call to discuss your project and whether we are a good fit. Available for a small number of new projects.
Or email → rohitsurvase.creates@gmail.com
AI · Fintech
AI fraud agents for banks (YC25), designed to work within the review queue investigators already trust.
Problem 01
An agent marking a cheque "suspicious" is only useful if the investigator can see why, and check it for themselves in a few seconds.
Every field the agent reads is boxed right where it found it on the cheque: green when it matches what that account normally looks like, red when it doesn't. Below it, each extracted value sits in its own chip — signature, amount, date, payee, cheque number, bank and routing number — so the one mismatch stands out at a glance, with the back of the cheque one tap away.
Problem 02
A loan filing pulls the same field from several documents, and they don't always agree. Quietly picking one value is exactly how a reporting error slips through.
Each field on the loan record shows where its value came from — the profile or the loan data — and which page. A field with conflicting values is flagged in place. Open it, and every candidate value is listed with how many documents it was found in, alongside a custom option and a note on why the chosen one is right, saved as a version for whoever examines the filing next.
Problem 03
Community development activity is reported by purpose. A purpose with nothing logged under it shouldn't quietly drop off the chart — that reads as missing, not zero.
Dollars and hours are both broken down by the same four purposes, and all four always stay on the chart, even at zero. A third chart splits logged activity into approved, awaiting review and rejected, because only approved activity counts toward the community development test.
Outcome
Socratix builds fraud agents for banks, coming out of YC25 — designed to fit into a review queue investigators already trust, instead of asking them to trust something new.
AI · Fintech · Compliance
Audit and compliance software for a YC-backed startup, with an AI agent that shows its reasoning at every step.
Designed with Studio Infer
Problem 01
One way in was a general AI assistant you could open from anywhere. The other lived inside each control, tied to a specific test. Both produced suggestions, but neither knew the other existed — the same request could land two different answers depending on which one you used.
Gave the control-level agent local authority. The general assistant stayed for open, cross-control questions — suggesting one action at a time to approve or skip, with its reasoning open alongside: the evidence it studied and the plan it made. Anything about one specific control now opens inside that control's own panel, next to the samples and test plan it's reasoning about.
Problem 02
Once agents were running across dozens of controls, a flat status like "in progress" could mean five different things depending on who you asked — running with the agent, waiting on a person, in review, or already signed off.
Gave every request exactly one of four explicit states — pending, in progress, in review, completed — and grouped the queue by them. A reviewer can filter straight to their own queue. On the controls sheet, each control sits against the principle it tests, with its results right there in the row.
Problem 03
A control's status showed up as a progress bar on the dashboard, a tag in the request queue, and a sign-off panel inside the control itself — three different visual languages for the same fact.
Kept the same four states recognizable everywhere — same colors, same language — so "waiting on a person" means the same thing no matter which screen you're on.
Problem 04
A preparer's approval and a reviewer's sign-off sat side by side, same size, same weight — exactly the setup that gets someone to sign off on work they were only supposed to draft.
Split them visually — one primary, one secondary — so which hat you were wearing was obvious before you clicked.
Outcome
Audrion shipped start to handover in eight weeks. The company was later acquired by Opendoor.
AI · Data analytics
AI data analytics that starts by understanding the data, so every answer rests on a solid foundation.
Designed with Studio Infer
Problem 01
Mapping relationships between entities is a judgment call — two columns that look alike but mean different things. Buried inside a progress bar, a bad guess there would flow silently into every report built on top of it.
Every entity the pipeline finds now carries its own confidence score, right on the card — low enough, and it's plainly asking to be checked before anyone builds an analysis on top of it.
Problem 02
A low number tells you to look. It doesn't tell you what you're looking at, or whether someone already decided this one was fine.
Opened a low-confidence relationship straight into what it's claiming and what it's called elsewhere — cardinality, source and target table, a linked glossary entry, and a tag for anything sensitive.
Problem 03
A single "ask anything" box has to guess what you actually mean — scoped to which data, answered by which logic.
Let people scope a question to specific entities and hand it to a named agent built for that kind of question, instead of hoping one general assistant gets it right.
Problem 04
Once the model of the data held up in one place, that same understanding had to show up consistently in chat, in agent comparisons, and on a canvas — or the trust built in one screen would leak out of the next.
Charts, comparisons and canvases all pull from the same underlying entities and confidence scores, so a number means the same thing wherever it appears.
Outcome
Product design for AI data analytics, built around the idea that understanding the data is the actual product — not a loading screen in front of the real one.
Mobility · Consumer
A ride app for auto-rickshaws that makes everyday trips effortless: set up a routine once, then ride on schedule.
Designed with Studio Infer
Flow 01
The home screen leads with one thing: how soon an auto can get to you. Everything else — saved places, coin offers, a nudge to turn a repeated trip into a routine — sits below the fold.
Confirming a ride shows the fare with coins already applied, waiting charges spelled out, and offers one tap away. Once a driver is assigned, the OTP, the driver's name and plate, and every stop on the trip stay on a single screen — with the fare and a way to edit the trip right there.
Flow 02
Where instant rides aren't available yet, the home screen changes what it offers instead of turning into a dead end.
The same screen points to the two things that still work there: renting an auto by the hour, or scheduling a recurring Auto-mate ride.
Flow 03
School runs and weekly classes are the same trip over and over. Auto-mate sets one up once — same pickup and drop, the same driver if you want one, and any day skippable without losing the slot.
Each routine shows its next pickup and driver preference. A routine's own page lists every upcoming ride, with completed, upcoming and cancelled counts. When someone keeps taking the same trip, the app offers to turn it into a routine.
Flow 04
For errands that hop between stops, renting an auto by the hour comes with the price and kilometres included laid out before you book.
One coin is one rupee off any ride, applied automatically at checkout. Coin and trip packs live on their own screen.
