What we work on

Our Research

Natural language processing, computer vision, machine learning foundations and large language models — the questions we work on, and why they matter.

How we work

Our research philosophy

We believe AI research should be accessible, ethical and impactful. Our work focuses on closing the gap for low-resource languages, particularly Bangla, while holding to the highest standards of scientific rigor and responsible development.

Focus on impact

We prioritize research that addresses real-world challenges and makes a difference to communities.

Rigorous methodology

Careful experimentation and validation, so that our results hold up.

Collaborative approach

We work with researchers and institutions around the world to move the field forward.

Innovation first

We encourage new approaches to hard problems in AI.

Area 01

Machine Learning Foundations

New optimization algorithms, meta-learning strategies and models that stay robust under distribution shift.

  • Optimization
  • Meta-learning
  • Robustness
  • Distribution shift

Open questions

  1. 1Training efficiency for large-scale models
  2. 2Generalization across diverse domains
  3. 3Handling distribution shift in real-world scenarios
  4. 4Theoretical understanding of model behavior

Area 02

Deep Learning Architectures

Efficient transformer variants, graph neural networks and multimodal representation learning.

  • Efficient transformers
  • Graph neural networks
  • Multimodal learning

Open questions

  1. 1Model scalability and computational efficiency
  2. 2Architectures for resource-constrained environments
  3. 3Cross-modal representation alignment
  4. 4Training stability for deep architectures

Area 03

Natural Language Processing

Language technology for low-resource languages such as Bangla — translation, alignment and prompt engineering.

  • Low-resource languages
  • Bangla NLP
  • Machine translation
  • Prompt engineering

Open questions

  1. 1Cross-lingual transfer for low-resource languages
  2. 2Prompt robustness and injection
  3. 3Alignment for diverse linguistic contexts
  4. 4Efficient fine-tuning for specific domains

Area 04

Computer Vision

Vision transformers and deep learning for medical and agricultural imaging.

  • Vision transformers
  • Medical imaging
  • Agricultural imaging

Open questions

  1. 1Domain adaptation for medical and agricultural images
  2. 2Limited labeled data in specialized domains
  3. 3Real-time inference for practical applications
  4. 4Robustness to varying imaging conditions

Area 05

Large Language Models

The capabilities, alignment and ethical considerations of large language models.

  • Alignment
  • Evaluation
  • Responsible AI

Open questions

  1. 1Understanding emergent behaviors in LLMs
  2. 2Mitigating biases and ensuring ethical use
  3. 3Scalable training and fine-tuning techniques
  4. 4Evaluating LLM performance across tasks

See the work itself

Every paper the lab has published, or get in touch about a collaboration.