Emovere
A proposal and one early prototype, June 2025. The prototype’s code was never pushed and is gone. The proposal survives.
Emovere is an offline voice for places with no connection, like displacement camps or a disaster zone. It keeps someone in acute distress company until a trained person can get to them.
What it is not
It is not a therapist and it doesn’t diagnose. It’s a first-aid bridge: something that talks with a person and doesn’t leave, for the minutes before help arrives. When someone needs a person, it says so and keeps talking until that person is there.
The first prototype
In June 2025, at 15, I built a working prototype on Gemma 3 4B that ran offline in 4 GB of graphics memory, and entered it in C# Corner’s AI for Peace challenge. That idea is what got me into IIT Kanpur’s AETL programme.
The proposal describes a faster pipeline: voice detection, then Whisper for speech, a small Llama model for the reply, then a local voice. It aims to start speaking in under 400 milliseconds, on a Jetson board or a consumer GPU, with no uplink at all.
The part I think matters
The pipeline is the least interesting layer. The real idea is a memory for emotion wrapped around a model that has none.
Each turn, an emotion classifier reads what the person said. Emovere doesn’t just react to that one reading. It keeps a running emotional state that carries over between turns and fades at a set rate, like momentum. That state then rewrites the model’s instructions for the next reply. So how it speaks depends on how the conversation has gone, not just on the last sentence.
None of this has been properly tested. There is a way to test it without knowing anyone’s true feelings: pick the fade rate that best predicts the classifier’s next reading on held-out conversations.
Why Gemma 3 4B is still its voice
Newer models are more capable but sound more like assistants. Gemma 3 4B still sounds like company. I wrote about that here.