TalkiEmma — trust layer
API for EdTech Integrators

The trust layer for AI systems built for children.

Is your AI assistant talking to students? Emma sits between your application and the AI model: she guides with hints without giving away the answers, redacts personal information, detects distress, and aligns with the curriculum. Whichever model you use, today and tomorrow.

Request sandbox access View Integration
Education-firstDesigned for learning: hints, never direct answers.
Child-safety alwaysPII, distress, toxic content: the firewall is deterministic, not optional.
Model-agnostic by designThe model is a replaceable component. Trust remains unchanged.
Integration

Two modes, the exact same guarantee.

Whether you are building from scratch or already running your own assistants, Emma integrates in hours.

Managed Mode

Emma responds for you

Send the conversation to POST /v1/chat — Emma handles the model, keys, inference costs, and runs the firewall on both sides. No LLM provider account needed. Image input supported.

Guardian Mode

Your assistants, our firewall

Keep your models and servers: send the student input and raw model output to POST /v1/guard. Emma returns a verdict and a safe, adapted version. Purely deterministic — no model calls.

Pedagogical Firewall

If the model fails, the firewall catches it.

Safety rules are not system prompt instructions that a model might bypass: they are deterministic validations applied to every single exchange.

Hints, never answersThe final solution is automatically detected and hidden, even if the model leaks it.
PII ScrubbingBi-directional masking of personal identifiable data before it reaches the model and before it reaches the child.
Distress DetectionImmediate alert sent to the integrator for human routing — with zero model dependency.
Curriculum GuardOut of curriculum topics trigger a polite redirect and refusal (out_of_curriculum verdict).
Anti-False-PraiseNo unearned praise: encouragement follows actual student effort and progress.
Prompt ShieldingJailbreaks and hidden system role manipulation attempts are stripped and neutralized.
Compliance

Designed to retain nothing it shouldn't.

European Union hosting. No conversation retention. Pseudonymous identifiers only — no names, e-mails, or schools stored in the API.

Emma applies a security framework aligned with ISO/IEC 27001, ISO/IEC 42001, GDPR, and the AI Act. Independent certification is planned.

Request sandbox access

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

Request sandbox access