TalkiEmma — trust layer
Trust Architecture

Policy Enforcement Layer for Educational AI.

Emma is not an LLM. It is a bi-directional, deterministic barrier that intercepts, validates, and transforms every turn between the student and the AI assistant.

A horizontal and agnostic architecture

Emma is strictly positioned horizontally between the client application and the inference model infrastructure (LiteLLM, AWS Bedrock, OpenAI). As a Policy Enforcement Layer, it enforces security guarantees that system prompts cannot guarantee 100% of the time.

Pre-Inference Filtering (Input) Every student message is scrutinized by the firewall_pre module. Personal Identifiable Information (PII) is redacted and prompt injection attempts (jailbreaks) are neutralized.
Post-Inference Filtering (Output) The model output is checked by the firewall_post module. If a direct solution is detected (answer leak), the response is blocked and rewritten to a child-friendly, age-appropriate hint.

Request sandbox access

Test Emma on your own product — sandbox is free of charge.

Request sandbox access