Introducing MMI Labs
Today we are starting MMI Labs, the research team at MM Intelligence.
For the past year we have been building AI for clients in Alabama and the Gulf. The same hard problems kept coming back. A model has to run on a network with no internet. A meeting switches between Arabic and English every few sentences. A voice agent has to answer a caller fast enough that the pause never feels strange. Each time we solved one of these, the answer lived in a client folder and in our heads. MMI Labs is where we start writing it down.
What we will publish
Our work runs along three lines, and every article here will sit on at least one of them.
AI that runs where your data lives. Some data cannot leave the building. We get capable language models running on hardware inside a customer's walls, including networks that are fully cut off from the outside. That covers the hardware, the model, how people reach it, and how it gets updated safely.
Speech that works in Arabic and English. Most speech tools were built for one language at a time. Gulf teams mix Arabic and English in the same sentence, and they speak dialects that most training data does not cover. This is the research behind Majlis, our meeting transcription app.
Voice agents that hold up on real calls. A voice agent is a chain of parts: the phone line, live transcription, the language model, and the voice that answers. Each part adds delay, and people interrupt. We write about how we keep the whole chain fast and steady.
How we will publish
- Real numbers when we have them. When a figure is an example, we say so on the figure itself.
- What failed, not only what worked. Most of what we learn comes from the first version that did not hold up.
- Nothing about client contracts. We describe the shape of the work in general terms. Customer names, sites and details stay private.
What comes next
Over the coming weeks we will publish one article at a time, starting with these three:
- Running a local language model on an air-gapped network
- Teaching a speech model to switch between Arabic and English
- Where the time goes in a voice agent
If you run into one of these problems in your own work, we would like to hear about it. Get in touch.
