Mihir Barve — Product Manager

The brief

SuperSales needed its Dialler to stop losing the 70% of calls that go unanswered. The ask was voicemail support. The opportunity was bigger than the ask.

The pain

A rep who hits voicemail has two options: hang up, or spend forty seconds leaving the same message for the hundredth time that day. Multiply by every unanswered call and entire afternoons disappear into answering machines.

The call

Automated voicemail drops existed. But they were one recording blasted to everyone, and callers can smell a broadcast. The call was to make the drop personal: AI-generated voicemail libraries, tailored per campaign, so the message that lands sounds like it was meant for that person. The rep taps drop, the system leaves the right message, the rep is already dialling the next lead. Personal at scale is what makes a voicemail earn the callback, and that is what was first to market.

The wireframesframe 01 / 03

Voicemail Settings (the library lives here) and the Add Voicemail modal. Note the states: empty library, mid-generation, per-campaign defaults. The unglamorous states are where engineers usually have to come back and ask. They did not.

Voicemail Settings and library wireframe

The library, living in Settings beside Email and Value Prop Templates. Every row plays inline, so a rep can audition a drop without leaving the page. Uploaded By is there because in a team of twenty, whose voice is on the recording matters.

What shippedframe 01 / 02

The shipped screens, beside the drawings.

Shipped. Both branches survived, and the build added what only a build reveals: a default name in the field, the accepted file types spelled out, and an Add button that stays disabled until there is something to add.

<900 → 4,087calls / month · Oct → Jan

The Dialler this feature belonged to went from under 900 calls in October to 4,087 by January. Roughly 5x call volume in three months. That number is call volume, nothing else, and it was a team result the Dialler carried.

The brief

Wingman (later Ruzo.ai) V0 was live. Real usage did what real usage does: exposed friction no spec could have predicted. V1 needed to fix exactly the right things without relitigating the whole product.

The pain

The usual revision cycle is a meeting, a doc nobody reads, and a JIRA backlog that loses the plot. Context dies between the person who saw the problem and the person who fixes it.

The call

Write the edits into the wireframes themselves. Each of these four frames carries its own critique: what was wrong, what changes, why. The artefact is the handoff. No meeting required, no context lost, nothing to misread.

The wireframesframe 01 / 04

All four annotated frames, legible. Read the margins; the margins are the work.

Sign In. Two edits in the margin: drop the placeholder Chat Bot name for Wingman, and give the dead left half an actual welcome. The critique sits on the frame, so nobody has to remember what was said in the meeting.

What shippedframe 01 / 02

V1 production screens beside the annotated frames. The distance between annotation and shipped pixel is the point.

Shipped sign-in. Both notes from the first frame landed: the placeholder name is gone and the empty half carries a welcome. The product was renamed twice between the drawing and this screen, which is what a pivot actually looks like.

The brief

Sales calls get handed off: SDR to AE, agent to specialist. The Dialler needed transfers that did not torch the conversation in the process.

The pain

A cold transfer is a dropped baton. The caller repeats everything, the receiver has no context, and the moment that made the handoff worth doing dies in the hold music.

The call

Warm transfer: the receiver gets briefed while the caller stays live, then the handoff completes. Which sounds simple until you draw it honestly. What if the receiver declines? Does not pick up? Is on another call? What does each party see at each second? This flow is mostly failure paths, and the wireframes treat the failure paths as the main design surface, not the fine print.

The wireframesframe 01 / 08

The full modal flow plus the transfer extension menus.

The entry point. The rep is live on a call and hits Transfer. The picker opens with users, teams and their extensions, and the three actions along the bottom are deliberately greyed out because nothing has been selected yet. Disabled is a state, and it gets drawn.

The brief

The AgentR pivot needed a complete agentic AI platform. Not a feature on an existing product: the whole product, from first principles, at pivot speed.

The pain

Zero-to-one design usually ships as a hero screen and a prayer, with every dropdown, modal, and edge state left as an exercise for the engineer. That is how V1s arrive six months late wearing a V0 badge.

The call

Draw everything. All twenty screens: sign-in, the chat surface with its artifact window, every dropdown in its open state, projects, knowledge and instruction modals, chat history, bulk actions, settings, billing. If an engineer would meet a state, the state exists in the flow. A month of build, drawn in an evening, no questions back.

The wireframesframe 01 / 09

Four clusters up close (chat + artifact window with states, projects, knowledge modal, billing) and the full twenty-screen contact sheet.

Chat with artifact window wireframe, state 1

Inside a chat with the nav collapsed to an icon rail. The task on screen is a real one, sending an email, because a chat surface drawn with filler text tells you nothing about how it behaves while a task is actually running.

Motor Racing, Auto Journalism
& Real Estate moulded me
into a Product Manager
— not a bootcamp.

Four careers. One way of thinking.

The work, shown working.

Two stories with everything on the line.

The AgentR Pivot.

The SuperAGI Dialler.

The paper version.

The part where we connect.