What I've built

AI Products, End-to-End

Three production-grade AI products I conceived, wrote the PRDs for, designed, and directed to build — AI-assisted, validated as working prototypes. Each follows the same arc: real customer problem → sharp product thinking → shipped software → measurable design targets.

Voice AI Platform

MB Voice — AI Voice-Calling Platform

Role: Product lead + build direction · AI-assisted

A self-hostable AI voice agent that answers inbound support and runs outbound sales calls in 12 Indian languages — built to help a leading B2B healthcare marketplace scale customer conversations without adding headcount.

LiveKitPythonSarvam Indic STT/TTSGemini / OpenAIFastAPINext.jsReact Flow

Problem

Human-only calling couldn't scale: leads went cold before a rep called back, repetitive support queries tied up agents, after-hours and multilingual demand went unanswered, and manual dialing risked TRAI/DLT compliance breaches.

Approach

  • Positioned AI as an additive layer — it qualifies, resolves, and routes; humans still close.
  • Four agent profiles: support hotline, order updates, payment reminders, lead qualification.
  • Compliance-first: calling-hours, consent, and DNC gating on every dial.

What I built

  • A no-code, drag-and-drop visual call-flow builder for non-engineers.
  • A lead → auto-qualify → CRM pipeline with intent, sentiment & BANT scoring.
  • Automated per-call QA + compliance scoring, and mid-call language switching.

Impact (design targets)

Engineered around sub-2-minute speed-to-call, a ≥70% connect rate, 100% of calls auto-QA'd, and ~₹5–6 per 3-minute call — validated end-to-end on seeded data.

12
Indian languages
~₹5–6
Target cost / call
100%
Calls auto-QA'd
4
Agent profiles

Prototype built and verified on seeded/mock data; metrics are design targets, not live production results. Employer specifics anonymised.

AI Sales CRM

SalesEdge — AI-Native Sales CRM

Role: Product owner + 15-week PRD + build direction

A CRM built on one belief: a CRM shouldn't just store data — it should drive sales behaviour. Designed to turn fragmented, slow, freestyle selling into a disciplined, system-enforced engine.

Next.js 16React 19Rules engineLLM-readyRecharts

Problem

Customer data lived across spreadsheets, WhatsApp, email and a static CRM. Leads were contacted after 24–78 hours, assigned randomly, and repeat buyers lost their owner — killing continuity and repeat revenue.

Approach

  • One source of truth for customer, lead, order & support data.
  • Rule-based scoring, auto-assignment, and mandatory next-actions.
  • Shared sales + support workflow with a clean AI/LLM upgrade path.

What I built

  • A prioritised Lead Engine with P1/P2/P3 scoring and ownership continuity.
  • Order lifecycle, cancellation/refund flows, and a support workspace.
  • Gamification and real-time manager dashboards.

Impact (design targets)

Targets lead-response time cut from 24–78 hours to 30–60 minutes, 95% ownership continuity for repeat buyers, 5+ tools consolidated into one, and AI call-prep in under 60 seconds.

24-78h → 30-60m
Lead response target
95%
Ownership continuity
5+
Tools replaced
<60s
AI call prep

Detailed 15-week PRD authored end-to-end; built as a working prototype. Metrics are design targets. Employer specifics anonymised.

Conversation Intelligence

CallEdge — Real-Time Call Guidance & Auto-QA

Role: Founder-style product lead + full build

Real-time call guidance and automated quality analysis, built India-first — every conversation guided while it happens, and understood after. A category proven abroad (Balto, Cresta), reimagined for Hinglish, WhatsApp, and CRM-native context.

Next.js 16React 19Real-time SSESarvam / Deepgram / ElevenLabs seamsMulti-tenantDPDP

Problem

Recorded calls were almost never reviewed — only a random 2–3% sampled. Coaching was anecdotal, compliance unmeasured, and a new multilingual export motion was launching with zero conversation tooling.

Approach

  • Empowerment-first: agents get instant, explained scores before managers get surveillance.
  • Hinglish and WhatsApp treated as first-class, not edge cases.
  • One engine, six surfaces — standalone SaaS + a CRM embed.

What I built

  • Live agent guidance with auto-ticking checklists and objection battlecards.
  • Auto-QA on 100% of calls against plain-English scorecards with evidence links.
  • A DPDP compliance centre (consent + PII redaction) and a no-code playbook designer.

Impact (design targets)

Moves QA coverage from ~2–3% sampled to 100% scored, auto-writes CRM summaries, and cuts after-call work — backed by a full 16-slide pitch deck, a 15-requirement PRD, and a 3-tier pricing model.

~2-3% → 100%
QA coverage
Real-time
Hinglish guidance
WhatsApp
First-class channel
DPDP
Compliance-ready

Full working prototype on a simulated Hinglish call corpus; metrics are design targets. Positioned against Balto/Cresta as market context, built as an original codebase.

Want the deeper story behind any of these?

I'm happy to walk through the PRDs, architecture decisions, and trade-offs — and how a decade of customer conversations shaped each one.

Let's talk →