The Tech Stack That Let a Digital Human Run a Live Scam on Real People

Immersive Tech

Pranay Bhandare

6mins

Mar 16, 2026

Financial fraud awareness has a fundamental execution problem. Brands publish guidelines. Banks send SMS alerts. Compliance teams run internal workshops. None of it simulates the actual pressure of a live scam call. People know the theory. When the moment arrives, they freeze.

HDFC Securities needed a different format entirely. Not a pamphlet. Not a webinar. A controlled, public encounter where real people faced a convincing digital persona running actual fraud tactics — in real time, in a premium mall, with a crowd watching.

This is a breakdown of how that system was built, what it demanded technically, and why the format produced results that static campaigns cannot replicate.

The Core Problem With Fraud Education

Most awareness content informs. It does not train.

There is a significant gap between knowing a scam exists and resisting one under pressure. That gap only closes through practice. The brain needs a rehearsal, not a reminder.

The brief demanded an installation that closed that gap in three minutes or less, ran continuously across three days, converted every physical interaction into a measurable digital touchpoint, and did all of this in a public space with zero tolerance for failure.

Digital Double: Building a Credible Persona

The avatar was modelled on actor Manoj Pahwa — a familiar face with an established sense of authority and familiarity in the target demographic.

Credibility was non-negotiable. A cartoonish or generic avatar would have broken the illusion immediately. The entire premise depended on participants genuinely feeling the social pressure of the scenario.

The team built a photoreal digital double from reference imagery. Skin texture, micro-expressions, facial geometry, and clothing were matched to the campaign reference with precision. Voice material from public sources was processed and shaped into a bespoke voice model. Intonation, cadence, and pacing were calibrated to feel natural and unscripted.

The result was a persona that read as real enough to trigger genuine hesitation — which was exactly the point.

The Interaction Architecture

Participants approached a kiosk anchored by a large transparent OLED display. A push-to-talk microphone gave users full conversational control. The setup mimicked the mechanics of a live phone call.

Five scam scenarios were available for selection. Each was drawn from documented fraud tactics:

OTP requests under false pretence
Requests for account or banking details
High-pressure urgency designed to force quick compliance
Persistent attempts to extract email addresses

Selecting a scenario altered the avatar's tone, escalation strategy, and conversational approach. No two sessions felt identical. The system was deliberately unpredictable.

Sessions ran for three minutes. Short enough to sustain attention. Long enough to build genuine pressure.

Scoring Resistance, Not Just Refusal

This was the design decision that separated the activation from a standard demo.

A passive "no" scored nothing. Creative resistance earned points. Participants who bluffed, stalled, asked verification questions, or redirected the conversation were rewarded. The conversational engine escalated pressure when users resisted — mirroring how real fraud calls actually behave.

An on-screen Aggression Meter tracked performance in real time. Crossing the 75% threshold declared the participant a winner.

Every session ended with a personalised profile summarising how the participant handled social pressure under realistic conditions. A QR code on-screen linked directly to the brand's online campaign — converting every physical interaction into a tracked digital entry point.

The Technical Pipeline

Building a convincing real-time digital human for a public environment is not a single-layer problem. Each component had to meet strict latency and reliability thresholds.

Digital Double Creation High-fidelity character build from multi-angle reference imagery. Priority placed on micro-expression fidelity and lighting consistency across display conditions.

Voice Synthesis Pipeline Cleaned audio samples processed into a responsive speech model. Intonation and cadence matched the reference persona. Output had to remain natural across repeated sessions without audible degradation.

Real-Time Conversational Engine Speech-to-text captured participant input. A context-aware response engine generated replies aligned with the selected scam scenario. Text-to-speech returned voiced responses. Lip sync and facial animation rendered simultaneously on the avatar.

Rendering and Hardware A high-performance GPU pipeline maintained sub-500ms round-trip times across the full interaction loop. That latency ceiling was essential. Anything above it collapsed the illusion.

Data Capture and Privacy No personal data was stored without explicit consent. Session-level metrics fed live dashboards only. Privacy compliance was built into the architecture from the start — not retrofitted after.

These layers operated as a single integrated system. Redundancy was built into critical modules to absorb load during peak footfall periods.

Results Across Three Days

The activation ran for 72 hours at a premium mall location.

2,000+ direct interactions
200,000+ impressions across live footfall and content amplification
100% session-to-digital conversion via on-screen QR
Every participant left with a personalised behavioural profile

Beyond the numbers, participants reported measurably higher confidence in identifying and resisting suspicious calls. The installation generated sustained conversation around fraud awareness that extended the campaign's effective lifespan well past the physical activation window.

Business Case: Three Specific Advantages

High recall per interaction. A short, intense session builds durable memory. Passive content does not. The intensity of a live scam encounter — even a simulated one — leaves a different cognitive imprint than reading a tip sheet.

Scalable scenario architecture. The underlying system does not need to be rebuilt to update content. Scam scenarios can be swapped, added, or localised without touching hardware. That makes the format reusable across markets, events, and regulatory cycles.

Direct funnel attribution. The QR-driven conversion model created a direct, measurable link between offline engagement and online campaign performance. Every interaction had a digital tail. Attribution was clean.

Where This Format Applies

The architecture described here is not specific to fraud prevention.

Any brand that needs to shift behaviour — not just raise awareness — can use this model. Risk communications. Compliance training. Product education at scale. Public health messaging. Any brief where the audience needs a practiced response rather than a passive one.

The format works because it places the participant in the situation before the situation finds them.

The Operational Principle

Keep sessions short. Prioritise latency above almost everything else. Make scoring meaningful and tied to the actual behaviour you want to reinforce. Ensure every interaction has a clear, frictionless digital follow-up path.

Those four principles are transferable to any behaviour-change activation running in a public environment.

About the Author

Pranay Bhandare
SEO Executive