Models and systems
We don't build models to be large. We build them to be good at a defined task — memory, speech recognition, voice generation, edge deployment. Each model is optimised to run independently with minimal infrastructure.
How we build
We take a principled path: start from the task, build small, and replace external dependencies with owned models as we prove each layer in production.
Principle
Each model in the Sankhya stack is designed for a specific task and optimized to be as compact as possible. We believe small models that do one thing well are more useful than large models that do everything adequately.
Strategy
We take a pragmatic path: use strong third-party APIs where they help us ship today, and replace the parts that matter most for cost, privacy, control, multilingual quality, and offline deployment with our own stack over time.
Direction
The end state is a stack of small models that can run independently — on-device, edge, or minimal cloud — serving learners and institutions without dependency on expensive GPU infrastructure or unreliable connectivity.
The stack
Each layer is designed for a defined task and optimised to be as compact as possible. Together, they form the teaching stack that powers SensAI.
Dhee
Learner memory layerAn in-house memory system that lets teaching agents carry context, preferences, and understanding across sessions. Dhee makes real personalization possible — not just prompt-level memory, but structured recall over time.
Akshar
Speech recognition for IndiaA speech-to-text stack being built for Indian accents, classroom noise, code-switching, and multilingual student interaction. Not a wrapper — a model trained on how India actually speaks.
Shlok
Instructional voice generationA text-to-speech model for natural, warm, Indian-language teaching voice. Built for the tone, cadence, and clarity that instructional speech demands — not generic assistant voice.
Compact Runtime
Edge deployment architectureA lightweight model architecture designed for low-end devices and poor-connectivity schools. The goal: the full Sankhya stack running independently, beyond cloud dependence.
Dhee
Learner memory layer
Dhee helps the teaching system remember how a learner has been taught, where they struggled, and what style worked better over time. Open-source on GitHub.
Akshar
Speech recognition
Being built for Indian accents, classroom noise, code-switching, and the realities of spoken educational interaction. Not a wrapper around existing STT.
Shlok
Speech generation
The voice layer being built for clear, natural instructional speech across Indian-language teaching scenarios. Warm, not robotic.
Compact Runtime
Edge deployment
A lightweight model architecture so the Sankhya stack can eventually run on lower-end devices and in weak-connectivity environments — independently.
Why this matters
Lower inference costs. Better privacy. Offline capability. Indian-language quality. Edge deployment. These are not features you can reliably get from general-purpose API providers — they require models you control.
Today
We use strong external APIs where they help us ship faster today — and replace the parts that matter most for quality, cost, privacy, and offline use with our own models over time.
What we own
Memory, voice quality, multilingual robustness, and edge deployment are the layers most worth controlling in-house. These are what differentiate the product.
What this unlocks
Small, focused models that we control means lower inference costs, better privacy, and the ability to deploy where connectivity is poor — the environments that need AI most.
