ECHO — AI Dynamic Rendering Engine

04 / 08

Featured project

GenAIPersonalisationGrowth

A real-time, intent-aware personalisation engine that rewrites landing page content based on search terms — making every paid acquisition page feel tailor-made for every visitor, at scale.

RoleProduct Manager (sole PM)

CompanyEmversity

Year2024

StackNext.js · Node.js · Redis · MongoDB · OpenAI · AWS Lambda · Gupshup

The Problem

The core problem

Emversity's paid acquisition was hitting a structural ceiling — not because of poor targeting or creative quality, but because of an intent mismatch at the moment of arrival. A student searching "operation theatre technician course Bengaluru" saw the exact same generic page as someone searching "nursing alternatives." The product experience treated every visitor as identical traffic.

The data confirmed what first principles suggested: visitor-to-lead conversion was stuck at ~2.1–2.2% against an industry benchmark of 5–8%. Microsoft Clarity session analysis revealed that 75% of users dropped off after the first scroll, spending an average of just 15 seconds on page — versus a 90-second industry norm. Cost per lead had climbed to ₹946 and D-7 ROAS sat at ~3.0, well below sustainable levels.

The reframe: this wasn't a marketing optimisation problem — it was a product relevance problem. The landing page was not a surface to tweak. It was an intent mismatch engine that needed to be replaced with a personalisation system.

The Solution

Building intelligence

ECHO was built as a three-layer intent preservation system — starting at acquisition, carried through activation, and handed off to the sales call. Guardrails built in: Semantic deviation rate capped at 0.82 cosine similarity threshold. Fallback to default content JSON when confidence is low, intent is ambiguous, or deviation is high. Graceful degradation at every layer — ECHO never breaks the page.

Layer 1 — Acquisition

Reads utm_term at page load. OpenAI rewrites hero headline, subheadline, and CTA in real time. Redis caches variants for sub-300ms delivery. Supports 15,000+ keyword variants.

Layer 2 — Activation

Intent-mapped WhatsApp templates sent post lead-capture via Gupshup. Human-in-loop approval. Open rate 55%→81%, reply rate 1%→8%.

Layer 3 — Sales Handoff

Intent summary pushed into LeadSquared CRM before first call. Sales agent sees the student's search context. First-call pickup 18%→30%, first-call conversion +112%.

Alternatives Considered

Form optimisation (optimising wrong metric)

→

Stronger incentives (solving symptoms)

→

Offer-led funnels (wrong sequencing)

Product Interface

←

→

ECHO — AI Dynamic Rendering Engine interface 1

ECHO — AI Dynamic Rendering Engine interface 2

Impact at a glance

Impact

+160%

Visitor → lead conversion

2.1% → 5.5%

+56%

First-call pickup rate

18% → 30%

−52%

Cost per lead

₹946 → ₹445

+112%

First-call conversion

to counselling

↓26pp

Bounce rate

72% → 48%

+137%

D-7 ROAS

from ~3.0 baseline

+2.2pts

Google Ads relevance score

6.1 → 8.3

+91%

D-15 ROAS

sustained improvement

My Role

End-to-end ownership

Sole PM end-to-end. Diagnosed the conversion plateau through behavioural data (Microsoft Clarity, CRM funnel analysis) and first-principles reframing. Personally interviewed 22 bounced users and audited the top 100 paid keywords. Defined ECHO's three-layer architecture and owned the trade-off to scope acquisition first, activation and sales handoff second. Led cross-functional execution with engineering, DevOps, marketing, and sales ops. Owned GenAI guardrails — semantic deviation threshold, fallback logic, and content boundary rules — to ensure ECHO never over-promised or hallucinated.

Key Learnings

What this taught me

Relevance over surface optimisation Most landing page conversion problems aren't UI problems — they're relevance problems. Lighthouse scores >85 and perfect CDN optimisation still left us at 2.1% conversion. The page was fast. It just wasn't relevant.

Intent is perishable Capturing it at acquisition means nothing if the next touchpoint — WhatsApp message or sales call — resets to generic. ECHO's value compounded because all three layers spoke the same language.

Hard guardrails > soft guidelines GenAI in a revenue-critical path requires hard guardrails, not soft guidelines. The semantic deviation threshold wasn't a nice-to-have — it was what made the system safe to run at scale across 15,000+ keyword variants.

Quality and volume move together The right metric to optimise for isn't leads — it's intent-qualified leads. Improving conversion from 2.1% to 5.5% while simultaneously improving D-7 ROAS +137% proved that quality and volume moved together when relevance was fixed.

Let's talk about this project ↗View all projects →

← Hercules ERP

samir.pm

Emerge → 42.ai →