You talk naturally.
My Hayat does the translation work internally.
Say it the way you would text it. The system figures out the rest.
What happens every time you send a message
Say it the way you'd text it
Arabic, English, Arabizi, or the mix you actually use. No special format required.
The system figures out what the moment needs
Safety, language, and context are processed before generation. Dialect failures and reasoning leaks are caught before you see the reply.
The reply is checked before you see it
A quality gate buffers output and retries on a different model if validation fails. Repeated failures fall back to conservative wording.
The conversation should lead somewhere
A useful question, a clearer pattern, a practical tool, or a reason to involve a human. Not an infinite loop.
What My Hayat actually is
Not fine-tuned on therapy transcripts. Not a replacement for a clinician. A carefully designed conversation system with deliberate constraints.
A Lebanese language layer
Arabizi recognition and generation is a separate engineering layer, not a model capability we assume. We maintain a curated Lebanese generation guide, separate from clinical knowledge, with explicit rules for what the model should and shouldn't generate.
Deterministic safety first
Crisis detection runs before the language model. It's a pattern-matching classifier, not a neural network hoping for the best. High-risk messages get a verified safe response — the generative model never controls a crisis response.
Evidence-informed, not fine-tuned
We use CBT, DBT and ACT frameworks as the basis for how the model should respond — through system instructions and a clinical knowledge retrieval layer, not by fine-tuning on patient data we don't have.
What My Hayat still gets wrong
This is a prototype. Honest limits are a feature, not a disclaimer.
Lebanese Arabizi generation is the hardest part. Generic foundation models default to Egyptian or MSA Arabic. Our guardrails catch many failures, but they can't automatically prove a sentence sounds native. Native-speaker review is ongoing.
The model sometimes misses the emotional intent behind a message and responds too clinically, or the other way around. We log these failures and they inform the next training pass.
Echoes — session memory and pattern tracking — is not live yet. Every conversation starts fresh. That's a privacy default, not a finished feature.
This is not emergency care. If you or someone you know is in immediate danger, call Embrace Lifeline: 1564.
What My Hayat currently does and what it doesn't claim
Conversations are not stored beyond the session by default
No account required for basic use
Safety classification runs before the language model
This is a prototype — privacy policy reflects actual implementation
