Under the hood.

How Ojas AI works, for technical teams evaluating it.

Model routing

Five tiers: fast local models (ONNX), Phi-4 14B, Qwen 3 8B, Sarvam for Indian languages, and a closed-API model only as a fallback. Each request goes to the smallest model that can handle it.

Restricted data

An India-only flag keeps processing in India; restricted data never goes to the fallback tier.

Memory and search

BGE-M3 embeddings in a Qdrant vector database power the organisational memory.

Knowledge (RAG)

Retrieval runs as its own application, separate from tool connections, so each is scaled and secured on its own.

Tool connections (MCP)

Nine internal MCP servers expose platform capabilities to agents as governed tools.

Agent-to-agent (A2A)

Agents pass work and information through a standard protocol, under the sharing rules.

Machine learning

Models are trained offline and deployed in ONNX format for fast, low-cost scoring and prediction.

Application stack

Node.js, MySQL, Redis and Angular in production.

Data residency

Everything the Ojas platform stores stays in India; built for the DPDP Act 2023.

Bring your technical team. We’ll walk them through it.

Talk to our team