Bhashini.ai Platform Architecture

Why We Go Deeper into the Architecture

As a platform that enables critical human–AI interactions across languages and use cases, Bhashini.ai considers it important to clearly explain how its systems are designed, how data flows through them, and how AI capabilities are delivered at scale.

The following architectural diagrams and technology disclosures are presented with the intent of demystifying our platform, establishing technical credibility, and enabling users, partners, and stakeholders to make informed decisions. By sharing our system context and internal container-level architecture, we aim to provide clarity on how different components interact, how security and scalability are addressed, and how services are responsibly operated.

Equally important, Bhashini.ai openly acknowledges the extraordinary contributions of the global open-source community that underpin many modern advances in language AI. Our platform thoughtfully integrates selected open-source models, datasets, and libraries across Text-to-Speech, Automatic Speech Recognition, Translation, and OCR, while layering on years of proprietary engineering, system integration, product development, and operational expertise to deliver secure, scalable, and production-grade services. We approach open source not as simple technology reuse, but as foundational building blocks that are used with transparency, proper attribution, and responsible innovation, and are strengthened by significant in-house engineering that transforms these foundations into real-world, deployable language technology solutions.

The sections that follow reflect this philosophy through transparent system architecture, clear attribution of open-source components, and a responsible presentation of how Bhashini.ai integrates these into its own production-grade platform.

Bhashini AI Platform - System Context

External Actors, Core Systems, and High-Level Interactions

This section explains who uses the Bhashini AI Platform, what systems exist, and how they interact, in a step-by-step narrative form.

1. People Involved

User
A person who uses Bhashini.ai services either from a web browser or a mobile application.
The user consumes AI services such as speech, text, translation, and e-books.

2. Bhashini Internal Systems

The Bhashini platform consists of three main internal systems:

2.1 Bhashini AI Services
This is the core AI backend system.
It provides the following capabilities:

Text-to-Speech (TTS)

Speech-to-Text (STT)

Language Translation

Optical Character Recognition (OCR)

These services are exposed using REST APIs and WebSocket APIs.

2.2 Bhashini.ai eBook Management System
This system is responsible for:

Managing e-books

Delivering DRM-protected e-books

It uses Readium LCP for digital rights management and content protection.

2.3 Bhashini.ai ERP System
This system manages:

User subscriptions

API credits

E-book purchases

Payments and billing

It is implemented using ERPNext / Frappe.

3. External Systems

The platform also interacts with external systems outside Bhashini:

3.1 B2B Client Applications or Agents
These are external applications built by Bhashini’s business customers.
They integrate with Bhashini AI Services using:

REST APIs

WebSocket APIs

Authentication is done using an API Key.

3.2 Sign in with Google
This is used for user authentication.
Google Sign-In works using the OpenID Connect protocol.

3.3 Razorpay Payment Gateway
This external system processes payments for:

Subscriptions

API credits

E-book purchases

It communicates with Bhashini.ai systems using secure HTTPS and webhooks.

4. Authentication Flow

The User authenticates using Google Sign-In.

Google verifies the user’s identity using OpenID Connect.

Once authenticated, the user can access Bhashini services.

5. AI Service Usage Flow

There are two main ways AI services are consumed:

5.1 Direct User Access
The user uses a web application or a mobile application to consume Bhashini.ai language technology services.
These web and mobile applications directly communicate with Bhashini AI Services using REST APIs and WebSocket APIs.

5.2 B2B Client Access
External B2B applications or agents also call Bhashini AI Services.

These calls are authenticated using API Keys.

The services provided are the same: TTS, STT, Translation, and OCR.

6. E-Book Access Flow

The user reads e-books using the Bhashini.ai eBook reading application on a web browser or a mobile device.

The eBook reading application communicates with the Bhashini.ai eBook Management System in the background.

This communication happens securely using HTTPS.

The e-books delivered to the user are DRM-protected using Readium LCP.

7. Payment and Subscription Flow

The payment flow happens in the following sequence:

The User purchases or renews subscriptions, credits, or e-books through the ERP system.

The ERP system initiates a payment request to Razorpay.

The User completes the payment on Razorpay’s platform.

Razorpay sends payment status updates back to the ERP system using webhooks.

After successful payment:

The ERP system sends subscription or credit information to Bhashini AI Services.

The ERP system sends e-book purchase information to the eBook Management System.

8. Overall Summary

The Bhashini AI Platform allows users and external business applications to consume AI language services such as speech, translation, OCR, and DRM-protected e-books. Users authenticate using Google Sign-In, make payments through Razorpay, and access services through secure APIs. An internal ERP system manages subscriptions, credits, and purchases, while the AI Services system delivers language intelligence and the eBook Management System handles protected digital book delivery. 

Bhashini AI Platform – Core Components

Internal Services, Data Stores, and Runtime Architecture

This section explains what runs inside the Bhashini AI Platform, what each part does, and how data flows between them.

1. Person Using the System

User
A user consumes Bhashini.ai language technology services such as speech, translation, OCR, and transliteration using:

A web browser, or

A mobile application

2. Overall System Boundary

All the components described below exist inside a single system called:

Bhashini AI Platform

This platform is internally divided into logical layers for clarity.

3. Edge Layer (Entry Point into the Platform)

The Edge Layer is responsible for handling user requests, security, and routing.

3.1 Reverse Proxy

Reverse Proxy

Technology: nginx (running directly on the operating system)

Responsibilities:

Serves static content such as:

Web application files

Encrypted e-books

Terminates SSL, meaning it handles HTTPS security

Load-balances incoming traffic

Routes API calls to the backend services

3.2 Web Application

Web App

Technology: Vue.js with Bootstrap CSS

Responsibilities:

Runs in the user’s web browser

Provides user interfaces for:

Text-to-Speech

Speech-to-Text

Translation

OCR

Transliteration

All backend communication is done securely through APIs over HTTPS

4. Mobile Application

Mobile App

Technology: Android with Kotlin

Responsibilities:

Runs natively on the user’s mobile phone

Allows users to:

Scan Indic documents

Convert text into speech

Consume language services in their native language

Sends all service requests through the same backend APIs as the web app

5. Core Backend Layer

Bhashini AI Backend

Technology:

Java

OpenLiberty

Docker containers

Responsibilities:

Implements all REST and WebSocket APIs for:

TTS

STT

Translation

OCR

Transliteration

Handles:

Authentication

Authorization

Acts as the central coordinator between applications, AI models, and data stores

6. AI and Data Layer

This layer contains all systems required for AI inference and data storage.

6.1 AI Inference Engine

Triton Inference Server

Technology: NVIDIA Triton Inference Server running in Docker

Responsibilities:

Executes AI and machine learning models

Supports both:

Proprietary models

Open-source models

Receives inference requests from the AI Backend using gRPC

6.2 Caching Layer

Redis Cache

Technology: Redis Cache (an extremely fast, in-memory data store) running in Docker

Responsibilities:

Caches frequently used inference results

Uses an LRU (Least Recently Used) eviction policy

Improves performance and reduces repeated computation

6.3 Persistent Data Store

CouchDB

Technology: CouchDB (a document-oriented NoSQL database) running on the host operating system

Responsibilities:

Stores:

Subscription information

API credit balances

Records:

Metered API usage

Billing-related logs

7. Request and Data Flow (Step-by-Step)

The user uses either:

A web browser, or

A mobile app

For web users:

The browser loads the web application and static content from the Reverse Proxy using HTTPS.

For mobile users:

The mobile app communicates with backend APIs via the Reverse Proxy.

The Reverse Proxy forwards API requests to the Bhashini AI Backend using REST or WebSocket protocols.

The AI Backend:

Validates user access

Checks subscription and credit details from CouchDB

Uses Redis to retrieve cached results if available

If no cached result exists:

The AI Backend sends an inference request to Triton Inference Server using gRPC.

The inference result is:

Returned to the AI Backend

Optionally cached in Redis

Logged for usage tracking in CouchDB

The final response is sent back to:

The web application or mobile app

And then presented to the user

8. Overall Summary

The Bhashini AI Platform consists of a reverse proxy, web and mobile applications, a Java-based backend, AI inference services, caching, and a database. Users access services through web or mobile apps. Requests pass through the reverse proxy to the backend, which manages security, billing, caching, and AI inference. AI models run on Triton Inference Server, results are cached in Redis, and usage data is stored in CouchDB. 

Bhashini.ai eBook Management Architecture

Secure Content Delivery, Licensing, and DRM Components

This description explains who is involved, what internal components exist, and how eBooks are published, purchased, licensed, and read within the Bhashini.ai eBook ecosystem.

1. People and External Systems Involved

1.1 Author or Publisher

An Author or Publisher creates and publishes eBooks on the Bhashini.ai platform.

They are later paid royalties for purchased books.

1.2 User

A User reads eBooks using the Bhashini.ai eBook Reader on:

A mobile device, or

A web browser.

1.3 Bhashini.ai ERP System

The ERP system manages:

eBook purchases

Payments

Royalties

It is implemented using ERPNext / Frappe.

2. Overall System Boundary

All components described below exist inside a single system called:

Bhashini.ai eBook Management System

This system is responsible for publishing, protecting, licensing, delivering, and tracking eBooks.

3. User-Facing Application

3.1 eBook Reader Application

Mobile App / Web App

Technology: Android (Kotlin) and Web

Responsibilities:

Allows users to read Indic language books

Supports books enriched with AI features

Uses DRM licenses to securely open encrypted books

The app does not store decryption keys permanently.

4. Core Backend Services

4.1 Bhashini.ai eBook Backend

Technology:

Java

OpenLiberty

Docker

Responsibilities:

Allows authors to publish eBooks

Manages:

eBook catalog

Users

Purchases

Coordinates license generation and delivery

Acts as the central controller of the eBook system

5. DRM and Encryption Components

5.1 LCP Encryption Tool

Technology: Go-based command-line tool

Responsibilities:

Encrypts newly published eBooks

Runs during the publishing workflow

Produces DRM-protected content

5.2 LCP License Server

Technology: Go, cryptography libraries, Docker

Responsibilities:

Generates LCP licenses for purchased eBooks

Each license defines:

Who can read the book

On which devices

Under what conditions

5.3 LCP Status Server

Technology: Go, cryptography libraries, Docker

Responsibilities:

Manages:

Device registration

Device activation

License revocation

Generates license status documents

Communicates with reading applications during book usage

6. Data Storage Components

6.1 Encrypted Content Repository

Technology: Network File System (NFS)

Responsibilities:

Stores encrypted eBook files

Contains no readable plaintext content

Books can only be decrypted using valid licenses

6.2 CouchDB Database

Technology: CouchDB (a document-oriented NoSQL database) running on the host system

Responsibilities:

Stores:

User purchase records

User-specific LCP license information

7. Publishing Flow (Author Perspective)

The Author publishes an eBook using the eBook Backend.

The backend invokes the LCP Encryption Tool to encrypt the eBook.

The encrypted eBook is sent to the LCP License Server.

The encrypted content is stored in the Encrypted Content Repository.

8. Purchase and Reading Flow (User Perspective)

The User purchases an eBook through the ERP system.

The ERP system sends purchase information to the eBook Backend.

The eBook Backend requests an LCP license from the LCP License Server.

The License Server notifies the LCP Status Server about the new license.

The eBook Backend stores the user’s license details in CouchDB.

The eBook Backend sends the LCP license to the user’s reading app.

The reading app downloads the encrypted eBook content from the content repository.

The reading app:

Retrieves license status documents

Registers the user’s device by communicating with the LCP Status Server.

9. Royalties Flow

The ERP system pays royalties to the Author or Publisher using UPI, based on eBook sales.

10. Overall Summary

The Bhashini.ai eBook Management System allows authors to publish encrypted eBooks and users to securely purchase and read them. Books are encrypted using LCP, licenses are generated and managed by dedicated license and status servers, and encrypted content is stored separately from licenses. Users read books through a mobile or web app using DRM licenses, while purchases, payments, and royalties are handled by the ERP system. 

Open-Source Technologies Used by Bhashini.ai

Beyond the Components Shown in the Architectural Diagrams

Why not build it yourself? Why Bhashini.ai?

Yes — today’s open-source ecosystem makes it possible to assemble ASR, TTS, and translation pipelines.

But building a demo is easy. Building a production-grade, real-time, multilingual voice AI platform is not.

Here’s what teams typically underestimate:

1. It’s not about one model — it’s about managing many

Real systems are not a single model call.

You need multiple components working together:

• ASR stack: speech recognition, VAD, language identification, speaker diarization
• TTS stack: text normalization, spectrogram encoder, decoder, speed control

Each component has:

• different latency profiles
• different failure modes
• different hardware requirements

Most DIY systems become fragile very quickly.

Bhashini.ai provides a unified orchestration layer that seamlessly manages multiple models in real time.

1. Real-time voice orchestration is non-trivial 

Handling live speech isn’t just running ASR.

You need:

• accurate speech start/stop detection
• pause/resume handling
• interim vs final transcripts
• seamless handoff to LLMs and TTS

Most teams underestimate this and end up with:

• laggy interactions
• broken conversational flow
• poor user experience

Bhashini.ai provides a battle-tested voice orchestration layer (VAD + FSM) that makes real-time voice interactions feel natural and responsive.

1. Latency and scalability are hard to get right

Real-world applications require:

• low-latency streaming
• high concurrency
• GPU-efficient inference

DIY systems often:

• break under scale
• become too expensive
• fail to meet real-time expectations

Bhashini.ai is built for real-time production workloads using optimized runtimes (ONNX, Triton, LibTorch, etc.) for each model.

1. Voice quality and consistency matter

Text-to-speech is not just intelligibility:

• naturalness, emotion, and consistency are critical
• voice cloning and cross-lingual voice preservation are non-trivial

Without proper processing:

• numbers, abbreviations, and formats sound wrong
• speech sounds robotic or inconsistent

Bhashini.ai provides 100+ high-quality, culturally relevant voices with advanced text normalization and speech control.

1. Enterprise-grade APIs are more than endpoints

Production systems require:

• authentication & authorization
• rate limiting and usage control
• persistent sessions
• reliability under load

Building and maintaining this stack (APIs + infra + scaling) is significant effort.

Bhashini.ai offers ready-to-use REST and WebSocket APIs built for enterprise-grade reliability and responsible AI usage.

1. Continuous improvement is expensive

Keeping up with:

• new models
• better datasets
• evolving benchmarks

requires ongoing investment and engineering effort.

Bhashini.ai continuously improves the stack — without additional effort from your team.

So why Bhashini.ai?

Because your goal is not to build language infrastructure — it’s to build your product.

With Bhashini.ai, you get:

• Production-ready multilingual voice AI platform
• Real-time orchestration across multiple models
• Low-latency, scalable infrastructure
• High-quality voices optimized for Indian languages
• Enterprise-grade APIs with built-in reliability
• Continuous upgrades without maintenance overhead

The bottom line:

You can build it yourself.

But you’ll spend months solving infrastructure, orchestration, and quality problems that Bhashini.ai has already solved.

Use Bhashini.ai — and focus on what makes your product unique.

Why Bhashini.ai? Why not BHASHINI.gov.in?

BHASHINI (by Government of India) is a powerful initiative to make Indian languages digitally accessible.

Bhashini.ai builds on similar goals — but is designed for a very different use case:

building real-time, production-grade applications.

1. Infrastructure vs Product Layer

BHASHINI.gov.in

• Focused on national-scale language infrastructure
• Provides access to datasets, models, and APIs
• Designed as a digital public good

Bhashini.ai

• Built as a production-ready platform
• Provides orchestration, optimization, and real-time capabilities
• Designed for application developers and enterprises

Think of it as:

BHASHINI.gov.in = foundation

Bhashini.ai = production layer on top

1. Models vs Complete Systems

BHASHINI.gov.in gives you access to:

• ASR, TTS, translation models
• datasets and benchmarks

But you still need to:

• integrate multiple models
• handle edge cases
• build pipelines

Bhashini.ai gives you a complete system:

• multi-model orchestration
• real-time voice handling (VAD + FSM)
• unified APIs

You don’t assemble — you integrate.

1. Real-time voice experience

Voice AI is not just model accuracy — it’s interaction quality.

BHASHINI.gov.in focuses on:

• model access and language enablement

Bhashini.ai focuses on:

• real-time streaming
• pause/resume handling
• interim vs final transcripts
• seamless conversational flow

This is critical for:

• voice assistants
• call automation
• conversational AI

1. Performance, latency, and scalability

Production applications require:

• low latency
• high concurrency
• efficient resource utilization

Bhashini.ai:

• uses optimized runtimes (ONNX, Triton, LibTorch, etc.)
• ensures consistent real-time performance

This level of optimization is essential for customer-facing systems

1. Voice quality and customization

BHASHINI.gov.in provides:

• strong foundational models and datasets

Bhashini.ai provides:

• 100+ curated voices
• text normalization (numbers, abbreviations, etc.)
• control over speech speed and delivery
• consistent, production-quality output

Built for brand identity and real user experience

1. Enterprise readiness

Production deployments require:

• authentication & authorization
• usage control and rate limiting
• reliability under load

Bhashini.ai provides enterprise-grade REST and WebSocket APIs ready for deployment.

1. Continuous evolution for applications

BHASHINI.gov.in evolves as:

• a national language ecosystem

Bhashini.ai evolves as:

• a product platform
• continuously improving orchestration, performance, and quality

So your application improves — without rework.

So when should you use what?

Use BHASHINI.gov.in if you want:

• access to datasets and base models
• research and experimentation
• language ecosystem participation

Use Bhashini.ai if you want:

• to build and ship real-world applications
• real-time voice experiences
• scalable, production-ready systems