Bangladesh may get its Grammarly. But what about our own ChatGPT?

Illustration: TBS

Bangla is often hailed as one of the sweetest languages in the world, but let's face it—writing it correctly is no walk in the park. 

From tangled grammar rules to tricky spellings, even seasoned writers sometimes stumble. Add to that the myriad regional dialects that differ from the standardised Bangla we mostly see in print, and it's no wonder crafting flawless Bangla text feels like a monumental task.

The problem has only worsened with the flood of incorrect Bangla content online. This influences AI tools like ChatGPT and Google's Gemini, often causing them to churn out erroneous Bangla.

Enter Shuddhi, a little-known language tool, accessible as a web application, Chrome extension, Firefox add-on and an Android app, which is trying to change the game. 

Developed by Sagor Sarker, the Lead NLP Engineer at Hishab Technologies Limited ( a computer and services company), Shuddhi is designed for everyday Bangla writing. It's not just another spellchecker—it fixes common grammar mistakes, corrects spelling errors and even helps you write Banglish (Bangla written in English letters) the proper way. 

And then, it also works as a translation tool, making it a versatile companion for anyone struggling to perfect their Bangla writing.

Bangladesh's energy and cloud infrastructure are currently not optimised for such high-demand AI workloads. Using cloud-based solutions also leads back to the funding challenge.

Faruk Ahmad, Senior AI Engineer at Deloitte in Tokyo, Japan.

Sagor believes Shuddhi has the potential to become a "Grammarly for Bangla," with a total of around 25 features, including paraphrasing, complete rewriting, plagiarism detection and fact-checking, in the pipeline, provided adequate funding and data training are secured.

But that's not all Hishab is cooking up. 

They have developed an LLM named "titulm-1b-bn-v1," which is trained with 1 billion parameters and 4.51 billion Bangla tokens, according to Mehedi Hasan Menon, another Lead NLP Engineer at Hishab. 

This bilingual LLM understands and generates text in both Bangla and English. Built using a decoder-style transformer architecture, it is well-suited for applications ranging from chatbots to translation support.

Additionally, they are experimenting with architectures such as Llama, MPT and Gemma and have developed Bangla LLM evaluation datasets, which they have also shared publicly.

Another key player in the Bangla AI scene is Ekush LLM, developed by Bangladeshi AI startup Intelsense AI through its platform Sense.ai. The company claims it to be "the nation's first large language model specifically trained in Bangla."

This advanced model leverages a transformer-based architecture to understand and generate Bangla text with high accuracy and fluency, enhancing applications such as chatbots, virtual assistants, content generation and sentiment analysis. 

Additionally, Ekush LLM incorporates proprietary Voice AI technology, enabling voice-based applications in sectors like telemedicine and agriculture. 

"Intelsense AI collaborates with local enterprises to co-develop domain-specific AI models and tools, aiming to improve workflow and productivity across various industries," said Sawradip Saha, Technical Lead at Intelsense AI. 

Through these collaborations, the company also aims to raise awareness among enterprises about the significance of AI and its potential to shape the future, which could ultimately help secure essential funding for the country's broader research community.

While these efforts are promising, many wonder if Bangladesh can realistically join the global AI race soon, especially when countries like the United States and China are dominating with GPT-4-level models to take control of the world's geopolitical, economic and technological futures. 

"The short answer is yes and no," said Faruk Ahmad, Senior AI Engineer at Deloitte in Tokyo, Japan.

He pointed out that developing smaller, domain-specific models is a more achievable goal—and some progress has already been made, as mentioned earlier. 

Companies are focusing on fine-tuning open-source models or building localised tools instead of starting from scratch. And so, "the scope of these models remains narrow and application-driven rather than being broadly general-purpose," shared Faruk.

Shafkat Khan Siam, an AI Engineer at Banglalink, shared similar views. "Most companies here rely on API-based models like OpenAI's GPT or Google's Gemini for their AI needs. While some fine-tune open-source models like Llama or Mistral for specific use cases, no one is innovating entirely new architectures."

And thus, replicating Bangladesh's very own foundational LLMs like GPT-4 or DeepSeek is nearly impossible for Bangladesh right now. 

The biggest reason is infrastructure—or the lack of it. 

Dr Farig Sadeque, Associate Professor of CSE at BRAC University, highlighted that training a large-scale LLM requires thousands of graphics processing units (GPUs), enormous datasets, and consistent power supply. 

For instance, OpenAI's GPT-3, with 175 billion parameters, was trained using approximately 10,000 NVIDIA A100 GPUs.

Also, the power consumed to train GPT-4 equalled the electricity usage of 1,000 US households for five to six years. 

"Bangladesh's energy and cloud infrastructure are currently not optimised for such high-demand AI workloads. Using cloud-based solutions also leads back to the funding challenge," Faruk said. 

And then there's the data problem. A major bottleneck is the lack of structured, high-quality Bangla datasets. Despite efforts to collect and generate NLP datasets, they still fall short of the scale needed for training a foundational model.

"While Bangladesh has a rich linguistic and cultural landscape, the availability of structured and high-quality datasets in Bangla and other local dialects is extremely limited. Unlike global tech giants, local companies lack the infrastructure and resources to collect, clean and annotate data at scale," Siam said. 

Data privacy is another concern. 

"Unlike the EU's GDPR (General Data Protection Regulation), Bangladesh currently lacks a robust data protection framework. As AI adoption grows, concerns about user data security and ethical AI usage will become increasingly relevant. Without clear legal frameworks, there could be ethical risks and resistance from users regarding how their data is collected and used in LLM training," Faruk said. 

With all things considered, it is highly likely that even if Bangladesh may soon get its own Grammarly-like language models for specific tasks, a full-fledged conversational AI like ChatGPT or DeepSeek to generate new content and provide extensive responses is still a long way off. 

But what is more important to understand is the necessity of developing LLMs in our country and whether it is really worth investing our resources into such an endeavour, said Dr Sadeque. 

According to him, when it comes to language coverage, GPT already understands Bangla more or less well, and it will continue to improve over time. However, where GPT falls short is in representing a country's cultural nuances. 

"To authentically capture a nation's culture through an LLM, it is essential to have deep insights into the people of that country and an understanding of its cultural nuances. Unfortunately, multilingual LLMs have yet to fully capture these nuances," he said.

For instance, if you ask an image generator to create a picture of a Bangladeshi man, it will most likely depict him with characteristics that are more Indian than a Bangladeshi with a distinctive Bangladeshi identity.

"Therefore, our efforts should now focus on ensuring that major LLMs are capable of capturing the cultural subtleties that define our identity," Dr Sadeque concluded.