Applied AI research studio, India
Applied AI,
built to ship.
Hypervect takes voice, language and vision models from paper to production for organisations that serve India. We measure before we claim, and we ship what we measure.
Our products
Two products, both in production.
SunAI
India doesn’t type. India talks.
A voice agent that talks with your customers and citizens in ten Indian languages, from your own material, and hands your team the result.
- Ten Indian languages, switched mid-call
- Replies in about two seconds
- Runs in India, and its data stays there
Hyperdex
The CRM your agents run.
One shared memory of every customer for your AI agents and bots. They log what happened, look things up in plain words and ask before they act.
- Free for everyone, including commercial use
- Works with any MCP agent or bot
- Each company’s data walled off, hosted in India
What we build
Four areas we build in.
Voice agents
Assistants that talk on the phone, in the browser and on WhatsApp.
Shipped as SunAI
Language systems
LLM applications grounded in your documents, websites and rules.
Shipped as Hyperdex
Vision and synthetic data
Detection, redaction and realistic synthetic datasets.
Data and evaluation
Custom datasets, annotation, and the testing that shows it works.
Selected work
A sewage treatment plant that answers its visitors.
A public awareness initiative at the Pirana Sewage Treatment Plant in Ahmedabad: a voice guide that visitors to the plant can talk to in their own language.
- 200million litres a day: the plant its visitors can now ask about
- Anyquestion, in Hindi, Gujarati, English or another Indian language
- QRcode at the plant: visitors scan it on their own phone
How we work
Three principles behind everything we ship.
Measured before claimed.
A number goes on a page only after we have timed it, judged it and read the transcripts. If we cannot measure it, we do not say it.
Built for how India speaks.
Indian languages first, code-mixing expected, the phone as a first-class channel. We design for the real caller, on a basic phone, in a busy queue.
Production, not prototypes.
We run what we build: live numbers, live links, live users. A pilot is a real deployment with a small scope.
In the lab now
Full-duplex Indian voice
An agent that listens while it speaks, so a caller can interrupt as naturally as they would a person.
Synthetic scenes for vision
Training detectors on rendered 3D situations that are rare, private or unsafe to collect.
Judging voice agents
Scripted test callers and model judges that score every language and model on the same questions.
Start a conversation
Bring us a problem that talks, reads or sees.
Tell us the job. We will say plainly whether we can do it, what we would measure, and what a first deployment looks like.