-
-
Bharat Setu
Inspiration
While interacting with local vendors and artisans across India, from bustling markets to offbeat rural regions, we noticed a recurring bottleneck. Despite having high-quality produce and crafts, these micro-entrepreneurs are restricted to their immediate physical geography. The barrier isn’t a lack of ambition—it is a lack of digital and linguistic accessibility.
Existing e-commerce and logistics apps mandate English or standard Hindi proficiency and assume high digital literacy. We built Bharat-Setu to dismantle this barrier. We wanted to create a platform that adapts to the user's native dialect, rather than forcing the user to adapt to the platform.
What it does
Bharat-Setu is a comprehensive, voice-first ecosystem featuring four core pillars:
Mandi-Connect: A hyper-local digital marketplace where vendors can list products entirely via voice commands in their regional dialects.
Parivahan: An AI-optimized logistics aggregator connecting small vendors with local transport (like mini-trucks and tractors) for affordable last-mile delivery.
Bhavishya: A predictive analytics module that forecasts market trends and prices to help vendors decide when to hold or sell.
Vani: A real-time voice translation assistant that sits between buyers and sellers, enabling a vendor speaking Bengali to seamlessly negotiate with a buyer speaking Marathi.
How we built it
We designed Bharat-Setu with a focus on high performance on low-end devices and scalable cloud infrastructure:
Frontend Interface: Built with React Native to ensure a smooth, cross-platform mobile experience from a single codebase.
Backend Services: Powered by Node.js and Express, connected to a highly available database to handle concurrent voice-data requests.
AI & NLP Engine: We utilized Amazon Bedrock for our core LLM tasks (intent recognition and dialect processing) and Amazon Polly / Transcribe for low-latency Speech-to-Text (STT) and Text-to-Speech (TTS).
Predictive Analytics (Bhavishya): We deployed machine learning models using Amazon SageMaker. To predict optimal selling prices, our algorithm factors in historical data, seasonality, and local supply volumes.
The core pricing prediction model minimizes the error in expected market value using a weighted regression approach. For any given commodity, the projected price $P$ at time $t$ is calculated as:
$$P(t) = P_{base} + \alpha S(t) + \beta D(t) - \gamma V(t) + \epsilon$$
Where:
$P_{base}$ is the historical baseline price.
$S(t)$ represents the seasonality index.
$D(t)$ is the real-time buyer demand score.
$V(t)$ is the current market volume (supply).
$\alpha, \beta, \gamma$ are weights optimized continuously via gradient descent.
For the Parivahan logistics module, we use a cost-minimization function to batch local deliveries together, reducing the individual freight burden for small vendors:
$$C_{total} = \min \sum_{i=1}^{n} (d_i \cdot r + w_i \cdot k)$$
(where $d_i$ is distance, $r$ is the transport rate, $w_i$ is the weight of the vendor's goods, and $k$ is a vehicle-specific constant).
Challenges we ran into
Code-Switching in NLP: Indian dialects frequently mix English words with regional languages (e.g., using "truck" or "rate" mixed into rural dialects). Standardizing these inputs for the LLM to recognize true intent was challenging. We had to refine our system prompts aggressively to handle unstructured, noisy audio data.
Voice Translation Latency: Achieving real-time negotiation requires translating STT -> LLM Translation -> TTS with minimal delay. We optimized our pipeline by moving to WebSockets for continuous bi-directional data streaming.
Data Sparsity: Gathering accurate, historical hyper-local Mandi data to train the Bhavishya model required significant web scraping and data normalization before we could feed it into SageMaker.
Accomplishments that we're proud of
We successfully built a working prototype where a user can open the app, speak naturally about what they want to sell, and the AI correctly parses the item, quantity, and expected price, instantly creating a digital storefront listing without a single typed word.
What we learned
Building Bharat-Setu pushed us to master integrating complex AWS AI infrastructure into a consumer-facing React Native application. We gained deep insights into prompt engineering for low-resource languages and learned how to optimize mobile architecture for areas with unpredictable network connectivity.
What's next for Bharat-Setu
Our immediate next step is to run a pilot launch in targeted regional clusters. We plan to expand the NLP models to support 15+ additional regional dialects and integrate native UPI escrow payments so vendors can receive secure, instant settlements directly within the app.
Log in or sign up for Devpost to join the conversation.