Building AI with MongoDB: How Devnagri Brings the Internet to 1.3 Billion People with Machine Translations

Mat Keep
January 23, 2024 | Updated: August 8, 2024
#genAI #Vector Search

It was while on a trip to Japan that Himanshu Sharma — later to become CEO of Devnagri — made an observation that drew parallels with his native India. Despite the majority of Japan’s population not speaking English, they were still well served by an internet that was largely based on the English language. Key to doing this was translation, and specifically the early days of automated machine translation. And so the idea to found Devnagri, India’s first AI-powered translation platform, was born.

“In India, 90% of the population are not fluent in English. That is close to 1.3 billion people. We wanted to bridge this gap to make it easy for non-English speakers to access the internet in their native languages. There are more than 22 Indian languages in use, but they represent just 0.1% of data on the internet,” says Sharma.

“We want to give people the same access to knowledge and education in their native languages so that they can be part of the digital ecosystem. We wanted to help businesses and the government reach real people who were not online because of the language barrier.”

Check out our AI resource page to learn more about building AI-powered apps with MongoDB.

Figure 1: Devnagri’s real time translation engine helps over 100 Indian brands connect with their customers over digital channels for the first time

Building India’s first machine translation platform

Sharma and his team at Devnagri have developed an AI-powered translation platform that can accept multiple file formats from different industry domains. Conceptually it is similar to Google Translate. Rather than a general consumer tool, it focuses on the four key industries that together make the largest impact on the everyday lives of Indian citizens: e-learning, banking, e-commerce, and media publishing. Devnagri provides API access to its platform and a plug-and-play solution for dynamically translating applications and websites.

As Sharma explains, “Our platform is built on our own custom transformer model based on the MarianNMT neural machine translation framework. We train on corpuses of content in documents, chunking them into sentences and storing them in MongoDB Atlas. We use in-context learning for training, which is further augmented with reinforcement learning from human feedback (RLHF) to further tune for precise accuracy.”

Sharma goes on to say, “We run on Google Vertex AI, which handles our MLops pipeline across both model training as well as inferencing. We use Google Tensor Processing Units (TPUs) to host our models so we can translate content — such as web pages, PDFs, documentation, web and mobile apps, images, and more — for users on the fly in real-time.”

While the custom transformer-based models have served the company well, recent advancements in off-the-shelf models is leading Devnagri’s engineers to switch. They are evaluating a move to OpenAI GPT-4 and the Llama-2-7b foundation models, fine-tuned with the past four years of machine translation data captured by Devnagri.

Why MongoDB? Flexibility and performance

MongoDB is used as the database platform for Devnagri’s machine translation models. For each sentence chunk, MongoDB stores the source English language version, the machine translation, and if applicable, the human-verified sentence translation.

As Sharma explains, “We use the sentences stored in MongoDB to train our models and support real-time inference. The flexibility of its document data model made MongoDB an ideal fit to store the diversity of structured and unstructured content and features our ML models translate.”

We also exploit MongoDB’s scalable distributed architecture. This allows our models to parallelize read and write requests across multiple nodes in the cloud, dramatically improving training and inference throughput. We get faster time to market with higher quality results by using MongoDB.
Himanshu Sharma, Devnagri co-founder and CEO

What's next?

Today Devnagri serves over 100 brands and several government agencies in India. The company has also joined MongoDB’s AI Innovators Program. The program provides its data science team with access to free Atlas credits to support further machine translation experiments and development, along with access to technical guidance and best practices.

Head over to our quick-start guide to get started with Atlas Vector Search today.

← Previous

Pledging Ourselves to the Future

As MongoDB’s sustainability manager, you could say I think about the climate a lot. After all, doing so is my job. But because it’s January and a time of reflection, I’ve been thinking about climate change more than usual — particularly about the progress we’ve made, but also the work that remains to be done. For example, in December the annual U.N. Climate Change Conference (COP 28) ended with a landmark agreement to transition away from fossil fuels, and the aim of reaching net zero carbon dioxide emissions by 2050. The COP 28 agreement also calls on countries to triple their renewable energy capacity and reduce other forms of emissions. The agreement was very welcome because before COP 28 began the U.N. released a stark report that showed national plans are, "insufficient to limit global temperature rise." As worried as I might be some days, I’m also buoyed by the climate action of the last few years. According to the U.S. Energy Information Administration, in 2022 more energy was generated by renewable sources than by coal for the first time. There have also been several regulations passed globally that make the measurement and disclosure of emissions mandatory, a key step in understanding — and reducing — emissions. MongoDB joins The Climate Pledge In the same spirit of optimism, I’m delighted to announce that MongoDB recently signed The Climate Pledge joining hundreds of leading organizations in publicly affirming our commitment to sustainability. The Climate Pledge’s hundreds of signatories commit to regularly report on their emissions and reach net-zero emissions by 2040 through decarbonization strategies and carbon offsets. “We’re thrilled to join the world’s leading companies — like MongoDB customers Verizon and Telefónica — in signing The Climate Pledge,” said MongoDB chief product officer, Sahir Azam. “MongoDB looks forward to working with the Climate Pledge team to ensure a more sustainable future for everyone.” Signing the The Climate Pledge is hardly the first step MongoDB has taken toward ensuring a more sustainable future. In 2023, MongoDB committed to being 100% powered by renewable energy by 2026, and achieving net-zero carbon emissions by 2030. To meet those targets, we’re working to reduce our carbon footprint through product innovation, by adding new sources of renewable energy, and by making MongoDB employees’ commutes more sustainable. Goodbye waste, hello (energy) savings In 2023, we also announced MongoDB’s new Sustainable Procurement Policy , which aims to ensure that sustainability is considered at all levels of our supply chain. The policy covers everything from the coffee we purchase (100% sustainably sourced) to the single-use items we use (restrictions leading to a 58% waste reduction in 2023). How MongoDB’s workloads are powered falls under our sustainable procurement efforts. Specifically, we’re currently working with our cloud partners — all of whom share MongoDB’s aim to be 100% powered by renewable energy by 2026 — to reduce our carbon footprint. "MongoDB takes its commitment to carbon reduction seriously, and we're fortunate to work with partners who share our enthusiasm for sustainability,” said MongoDB Lead Performance Engineer Ger Hartnett. “We look forward to continuing to collaborate with our partners on groundbreaking, energy-saving technology that makes real reductions in our carbon intensity." To meet our renewable energy target, we’ve focused our efforts on several areas, such as preferring buildings with renewable energy contracts or on-site solar when considering new office space. We’ve also entered into several virtual purchase power agreements (VPPAs). Virtual purchase power agreements are a great way for companies like MongoDB to invest in renewable energy without building anything on-site and are a proven method of adding renewable energy to the grid. Since 2022, MongoDB has worked with the enterprise sustainability platform Watershed to support renewable energy projects through VPPAs. Our first project helped build a solar plant in Texas that Watershed notes, “will avoid 13,000 tons of CO2, equivalent to taking nearly 3,000 gas-powered cars off the road each year.” And MongoDB recently signed a new VPPA that will support the development of solar panels for a factory in India. Solar energy is currently responsible for about 16% of global renewable energy, and only about 3.4% of overall energy in the U.S. Those numbers are sure to change, however. In the last fifteen years, global solar power generation has grown from 11.36 terawatt-hours to 1289.87 terawatt hours. What’s more, coal accounts for about 70% of India’s power — versus 20% in the United States — so projects like this will help reduce emissions across Asia. And because many MongoDB employees are directly impacted by air pollution in India , we see VPPAs as a way of benefitting the health and well-being of our employees, as well as the planet. MongoDB's stubborn optimism In the early months of the pandemic, Tom Rivett-Carnac, founding partner of Global Optimism — which launched The Climate Pledge with Amazon in 2019 — shared a video about shifting one’s mindset and changing the world . In the face of larger-than-life problems (like climate change), “stubborn optimism,” he said, “animates action, and infuses it with meaning.” “When the optimism leads to a determined action, then they can become self-sustaining … the two together can transform an entire issue and change the world,” he noted. “Stubborn optimism can fill our lives with meaning and purpose.” Composting is an example of a stubbornly optimistic action that’s both easy to adopt and one that (if enough of us do it) can change the world. Food waste accounts for 6% of global greenhouse emissions, and composting can help reduce those emissions. To put food waste emissions in perspective, 6% of global greenhouse emissions is roughly three times higher than annual global aviation emissions. In 2023, we also began tracking MongoDB’s waste and landfill diversion, and we’re working to improve how we dispose of waste by adding composting services to MongoDB’s hub offices. More than 80% of MongoDB’s offices already have composting services, and we aim to hit 100% in 2024. Not only have composting and single-use purchase reduction helped to decrease waste emissions, but both are highly visible to MongoDB employees. MongoDB employees are increasingly excited about sustainability, inspiring the creation of a mini-garden in our New York office, and the use of more sustainable commuting methods like biking. Though I tend to bike more for exercise than commuting these days (I’ve racked up more than 1,000 miles on my bike pass!), more and more MongoDB team members get to work in sustainable ways. For example, we’re rolling out electric vehicle commuting in India, an e-bike program was recently introduced in our Dublin office, and the bike locker in MongoDB’s New York HQ is generally packed. “I love biking to the office,” said Perry Taylor, a New York-based Information Technology Lead at MongoDB. “In addition to being a great way to stay fit, it’s awesome that how I commute helps the environment.” Looking back on 2023, I’m pleased with how much we accomplished toward MongoDB’s sustainability goals. At the same time, I recognize that more needs to be done. MongoDB enters 2024 with a renewed commitment to sustainability, and we look forward to furthering our progress. To learn more about MongoDB’s sustainability progress, please check out our Sustainability webpage and our latest Corporate Sustainability Report . For more information about fellow Climate Pledge signatories and an interactive timeline of progress made, visit The Climate Pledge .

January 23, 2024

Next →

AI-Powered Call Centers: A New Era of Customer Service

Customer satisfaction is critical for insurance companies. Studies have shown that companies with superior customer experiences consistently outperform their peers. In fact, McKinsey found that life and property/casualty insurers with superior customer experiences saw a significant 20% and 65% increase in Total Shareholder Return , respectively, over five years. A satisfied customer is a loyal customer. They are 80% more likely to renew their policies, directly contributing to sustainable growth. However, one major challenge faced by many insurance companies is the inefficiency of their call centers. Agents often struggle to quickly locate and deliver accurate information to customers, leading to frustration and dissatisfaction. This article explores how Dataworkz and MongoDB can transform call center operations. By converting call recordings into searchable vectors (numerical representations of data points in a multi-dimensional space), businesses can quickly access relevant information and improve customer service. We'll dig into how the integration of Amazon Transcribe, Cohere, and MongoDB Atlas Vector Search—as well as Dataworkz's RAG-as-a-service platform— is achieving this transformation. From call recordings to vectors: A data-driven approach Customer service interactions are goldmines of valuable insights. By analyzing call recordings, we can identify successful resolution strategies and uncover frequently asked questions. In turn, by making this information—which is often buried in audio files— accessible to agents, they can give customers faster and more accurate assistance. However, the vast volume and unstructured nature of these audio files make it challenging to extract actionable information efficiently. To address this challenge, we propose a pipeline that leverages AI and analytics to transform raw audio recordings into vectors as shown in Figure 1: Storage of raw audio files: Past call recordings are stored in their original audio format Processing of the audio files with AI and analytics services (such as Amazon Transcribe Call Analytics ): speech-to-text conversion, summarization of content, and vectorization Storage of vectors and metadata: The generated vectors and associated metadata (e.g., call timestamps, agent information) are stored in an operational data store Figure 1: Customer service call insight extraction and vectorization flow Once the data is stored in vector format within the operational data store, it becomes accessible for real-time applications. This data can be consumed directly through vector search or integrated into a retrieval-augmented generation (RAG) architecture, a technique that combines the capabilities of large language models (LLMs) with external knowledge sources to generate more accurate and informative outputs. Introducing Dataworkz: Simplifying RAG implementation Building RAG pipelines can be cumbersome and time-consuming for developers who must learn yet another stack of technologies. Especially in this initial phase, where companies want to experiment and move fast, it is essential to leverage tools that allow us to abstract complexity and don’t require deep knowledge of each component in order to experiment with and realize the benefits of RAG quickly. Dataworkz offers a powerful and composable RAG-as-a-service platform that streamlines the process of building RAG applications for enterprises. To operationalize RAG effectively, organizations need to master five key capabilities: ETL for LLMs: Dataworkz connects with diverse data sources and formats, transforming the data to make it ready for consumption by generative AI applications. Indexing: The platform breaks down data into smaller chunks and creates embeddings that capture semantics, storing them in a vector database. Retrieval: Dataworkz ensures the retrieval of accurate information in response to user queries, a critical part of the RAG process. Synthesis: The retrieved information is then used to build the context for a foundational model, generating responses grounded in reality. Monitoring: With many moving parts in the RAG system, Dataworkz provides robust monitoring capabilities essential for production use cases. Dataworkz's intuitive point-and-click interface (as seen in Video 1) simplifies RAG implementation, allowing enterprises to quickly operationalize AI applications. The platform offers flexibility and choice in data connectors, embedding models, vector stores, and language models. Additionally, tools like A/B testing ensure the quality and reliability of generated responses. This combination of ease of use, optionality, and quality assurance is a key tenet of Dataworkz's "RAG as a Service" offering. Diving deeper: System architecture and functionalities Now that we’ve looked at the components of the pre-processing pipeline, let’s explore the proposed real-time system architecture in detail. It comprises the following modules and functions (see Figure 2): Amazon Transcribe , which receives the audio coming from the customer’s phone and converts it into text. Cohere ’s embedding model, served through Amazon Bedrock , vectorizes the text coming from Transcribe. MongoDB Atlas Vector Search receives the query vector and returns a document that contains the most semantically similar FAQ in the database. Figure 2: System architecture and modules Here are a couple of FAQs we used for the demo: Q: “Can you explain the different types of coverage available for my home insurance?” A: “Home insurance typically includes coverage for the structure of your home, your personal belongings, liability protection, and additional living expenses in case you need to temporarily relocate. I can provide more detailed information on each type if you'd like.” Q: “What is the process for adding a new driver to my auto insurance policy?" A: “To add a new driver to your auto insurance policy, I'll need some details about the driver, such as their name, date of birth, and driver's license number. We can add them to your policy over the phone, or you can do it through our online portal.” Note that the question is reported just for reference, and it’s not used for retrieval. The actual question is provided by the user through the voice interface and then matched in real-time with the answers in the database using Vector Search. This information is finally presented to the customer service operator in text form (see Fig. 3). The proposed architecture is simple but very powerful, easy to implement, and effective. Moreover, it can serve as a foundation for more advanced use cases that require complex interactions, such as agentic workflows , and iterative and multi-step processes that combine LLMs and hybrid search to complete sophisticated tasks. Figure 3: App interface, displaying what has been asked by the customer (left) and how the information is presented to the customer service operator (right) This solution not only impacts human operator workflows but can also underpin chatbots and voicebots, enabling them to provide more relevant and contextual customer responses. Building a better future for customer service By seamlessly integrating analytical and operational data streams, insurance companies can significantly enhance both operational efficiency and customer satisfaction. Our system empowers businesses to optimize staffing, accelerate inquiry resolution, and deliver superior customer service through data-driven, real-time insights. To embark on your own customer service transformation, explore our GitHub repository and take advantage of the Dataworkz free tier .

November 27, 2024