Vedhak: A Server-Controlled Generative Simulation System for Adaptive Fraud-Resilience Training in Multilingual, Elderly Indian Populations

A networked system and method conducts an adaptive, multilingual, voice-based fraud-simulation training session between a human trainee and a server-controlled generative language model. On each conversational turn the model is invoked through a structured-output interface that returns both an in-character adversary utterance and machine-readable control signals comprising an inferred trainee state, a trust-value delta, and a discrete distress level. A server-side state machine applies the control signals to update a bounded trust value and advance the conversation through defined phases (open, rapport, extraction, closeout), assembling each subsequent prompt from the resulting state, thereby steering the open-ended model through a bounded pedagogical arc. Safety counters override the phase logic to cap simulated distress for a vulnerable trainee population. A resilience score is computed exclusively on the server from authoritative persisted interaction data with an early-exit bonus, retry-decay weighting, and a replay-decay factor, making the metric tamper-resistant. Voice is synthesised and recognised in server memory with no retention of biometric audio. A senior-initiated family-link mechanism permits a monitoring relationship to be granted only by the protected user and not seized by any third party. The system produces a measurable, longitudinal behavioural resilience metric constituting a concrete technical effect.