NOVA talks, and the screen keeps up

I led product strategy for a voice-first conversational AI platform at a GenAI services company in India. Recognizing that Indian enterprises prefer turnkey solutions over self-serve tools, I helped develop a dual product approach: an enterprise-ready voice platform (integrating STT, LLM, and TTS with features like RAG, turn detection, note-taking, and MCP) alongside an “Experience Center” - a website-embedded demo that showcased our capabilities and drove client conversions.


Through strategic partnerships (Deepgram, ElevenLabs, etc.) and competitive feature planning, we positioned the platform for India’s customer service market and accelerated go-to-market execution.

Experience Center

https://ai.think41.com/login

Role :

Designer

Product Strategy

Team :

2 Engineers

Timeline :

Jan 2025 -

Mar 2025

Skills :

Product Design & Strategy

Voice & Multimodal UX

Context

Trust doesn't come from a page

The starting point was a working loop, speech-to-text, a language model, and text-to-speech, grounded with RAG. Technically sound and not yet a conversation. It talked over people. It went quiet at the wrong moments.

Getting from a working pipeline to something that felt like talking to someone was the actual design problem, turn detection, preemptive responses, stronger guardrails, more providers, more languages, and eventually a way for the screen to move with it. The Experience Center, live today, is the fullest place where all of that shows up at once.

Getting from a working pipeline to something that felt like talking to someone was the actual design problem, turn detection, preemptive responses, stronger guardrails, more providers, more languages, and eventually a way for the screen to move with it. The Experience Center, live today, is the fullest place where all of that shows up at once.

5 +

5 +

personalized conversation paths, chosen by the visitor, not scripted

25

25

simultaneous voice conversations load-tested, zero incomplete sessions

<1s

<1s

average agent response time, mid-conversation

Research & Insights

Six ways to lose someone in under a minute

None of these came from a spec. Each one showed up the same way, I'd put the raw pipeline in front of someone, watch where the conversation actually broke, and design the fix.

None of these came from a spec. Each one showed up the same way, I'd put the raw pipeline in front of someone, watch where the conversation actually broke, and design the fix.

But Indian enterprises prefer turnkey solutions over self-serve tools.


That line justified building a second thing alongside the platform, not instead of it: the Experience Center, the same platform turned outward so anyone could see it work instead of being asked to trust a features list.

But Indian enterprises prefer turnkey solutions over self-serve tools.


That line justified building a second thing alongside the platform, not instead of it: the Experience Center, the same platform turned outward so anyone could see it work instead of being asked to trust a features list.

Solution

One capability at a time

Each gap from Research became a real product decision, not a backend toggle.


From prioritising the features based on the client demand and designing the interfaces to making strategic partnerships for voice AI providers (Deepgram, Elevenlabs, etc) I was actively involved in the making of this platform.

Each gap from Research became a real product decision, not a backend toggle.


From prioritising the features based on the client demand and designing the interfaces to making strategic partnerships for voice AI providers (Deepgram, Elevenlabs, etc) I was actively involved in the making of this platform.

CAI Platform is an enterprise-ready system to create, manage, and scale AI agents with ease.

The platform has to prove itself, not just describe itself . Voice and screen are one decision, not two, Silence is handled, not just speech. One provider or one language was never going to be enough. Safety isn't optional the moment a stranger is talking to it. The same rails have to hold a stranger's first conversation and the fiftieth one that day.

Features

What shipped

The Experience Center is a voice-first gateway into Think41, guided by Nova, our conversational AI. Instead of static decks, visitors explore the company through natural conversations enhanced with visual slides, interactive bento menus, and live demos.

User Flow

Sketching the conversation-first flow: how the UI frames intent, suggests next steps, and keeps the AI and user in sync.

These are a few concepts we explored for the Experience Center - how it should work, how GenAI could elevate the flow, and how the interface could support an ongoing AI-led conversation. You’ll see different ways we tried to structure intent, guide discovery, and make the assistant feel both helpful and in-control.


From sketchy ideas to lo‑fi wires, the goal was simple: design an interface that lets the AI do the heavy lifting while keeping the user confidently in the loop.

Designing Conversational Flow (Prompt Engineering)

To ensure Nova delivered consistent, human-like experiences, I designed structured system instructions covering tone, behavior, and guardrails. This prompt framework balanced brand identity with conversational flow.

How it works

  • A voice-first showcase of Think41, where users interact with Nova to explore offerings through conversational guidance + visuals.

  • Powered entirely by the CAI platform, which provides natural voice, context-aware nudges, and guardrails for human-like interactions.

  • Visual triggers (slides, demos, menus) are driven by CAI, making the experience dynamic, interactive, and multi-modal.

Responsive layouts for the screens to show context aware visuals during the conversations with AI.

Impact

Results of voice way to interact with websites

May 2025 - Aug 2025 data (Analytics from Posthog)

The dual product strategy delivered measurable business outcomes - from increased engagement to qualified pipeline growth. Bounce rate of the landing page also reduced by 11%.


  • Experience-led strategy improved website conversion and reduced bounce rates

  • Partnership approach enabled best-in-class voice A without building everything in-house

  • Dual product strategy solved both technology and adoption challenges for Indian market

Learnings

Ease of use is what creates impact

  • Market-first strategy: Understanding cultural preferences (turnkey vs DIY) shaped the entire product approach

  • Experience > Specs: In emerging tech, letting users try the product converts better than feature lists

  • Partnership velocity: Integrating best-in-class partners accelerated GTM vs building everything

"The best AI is the one you forget is AI - it just feels like conversation."