🧠 Research area
Bangla AI
Language, speech and multimodal models built from the ground up for Bangla and other low-resource settings.
Collaborate with us
What we work on
Sangbad-SLM & Bangla Language Models
Our 0.8B-parameter Bangla news language model and the family of compact models that follow it, trained from scratch on our own corpus.
Bangla Corpus Engine
Large-scale collection, de-duplication and cleaning of Bangla text, speech and dialect data: the foundation for every model we train.
Low-Resource Training Methods
Pretraining and fine-tuning techniques optimised for limited compute, so strong models stay affordable to build and run.
Speech & Dialect Research
Speech recognition and synthesis for Bangla and its regional dialects, built on field-collected audio.
Sovereign Benchmarks
Evaluation suites written for local contexts such as government, healthcare and education, so model quality can be trusted and compared.
Efficient Inference
Quantisation, distillation and serving research so our models run on modest GPUs and CPUs at low cost.
Research that starts with Bangla
The lab is organised around one flagship model and four research pipelines focused on Bangla-native, low-resource and sovereign AI — so every system we put into the world is one we understand end to end.
Sangbad-SLM
0.8B parameters · Bangla news language model
Bangladesh's first purpose-built Bangla news language model, trained from the ground up on our own Bangla news corpus. Designed to run on modest hardware with low inference costs, it outperforms larger general-purpose models on Bangla news comprehension.
- ✓ Trained on a proprietary Bangla news corpus
- ✓ Runs on-premise on modest GPU/CPU hardware
- ✓ Low inference cost for national-scale deployment
Bangla Corpus Engine
Large-scale collection and cleaning of Bangla text, speech and dialect data — the foundation for every model we train.
Low-Resource LLM Training
Fine-tuning and pretraining techniques optimised for limited compute, so strong models stay affordable to build and run.
Sovereign Benchmarks
Evaluation suites built for local contexts — government, healthcare and education — so you can trust the numbers.
National Inference Pipeline
An on-shore infrastructure blueprint for public institutions: your models, your data, your servers.
Founded in Dhaka as an AI research lab
AWS Activate grant recipient (2023)
Backed by BYLC Ventures and a Bangladesh accelerator
How we work
Frame
Problem statement, data audit, evaluation plan.
Research
Experiments, training and benchmarks against baselines.
Pilot
A time-boxed prototype with real users and real data.
Transfer
Deployment on your infrastructure, monitoring and training.
Interested in Bangla AI?
Tell us about your research question, your data or the people you serve, and the lab will get back to you within one business day.
Talk to the labOther research areas
Applied AI Research
Turning research into working systems for documents, vision, data and decision-making.
The lab's applied group takes models from the bench into the field: document intelligence for Bangla and English paperwork, computer vision, forecasting and retrieval-augmented assistants, each measured against real-world data.
- Document & Data Intelligence
- Computer Vision
- Predictive Modelling
- Retrieval & Generative Systems
- AI Evaluation & Safety
- Experimental Prototypes
Sovereign AI Infrastructure
On-shore, private infrastructure so institutions can run their own models on their own data.
Owning a model is only useful if you can run it. We research and build the inference pipelines, deployment blueprints and governance needed to operate AI on your own servers or private cloud, inside your borders.
- National Inference Pipeline
- On-Premise & Private Cloud
- MLOps & Model Monitoring
- Data Governance & Privacy
- Compute Strategy
- Open Standards & Interoperability
Innovation Programs & Partnerships
Joint research, pilots and capacity building with government, academia, NGOs and industry.
We work best alongside partners who bring a hard problem and real data. Our programs range from short pilots and proofs of concept to multi-year joint research, with training so partner teams can carry the work forward themselves.
- Joint Research Projects
- Pilots & Proofs of Concept
- AI Readiness & Strategy
- Capacity Building
- Data Partnerships
- Long-Term Support