Frequently asked questions
Straight answers about what the lab researches, how we partner with institutions and enterprises, and what to expect.
Hackules is an AI research and innovation lab headquartered in Dhaka, Bangladesh, founded in 2021. We build Bangla-native and low-resource AI: Sangbad-SLM and the corpora and benchmarks behind it, applied AI systems, and the sovereign infrastructure institutions need to run AI on their own terms.
Four research areas: Bangla AI (language, speech and multimodal models trained from scratch for Bangla), Applied AI Research (document intelligence, computer vision, predictive modelling and retrieval-augmented systems), Sovereign AI Infrastructure (on-premise and private-cloud deployment, inference pipelines, MLOps and governance) and Innovation Programs & Partnerships (joint research, pilots and capacity building).
Sangbad-SLM is our flagship model: a 0.8-billion-parameter Bangla news language model trained from the ground up on a Bangla news corpus we collected and cleaned ourselves. It is designed to run on modest GPU or CPU hardware with low inference cost, and it outperforms larger general-purpose models on Bangla news comprehension.
Bangla is where we started and where we have the deepest expertise, but the methods we develop for low-resource languages, limited compute and sovereign deployment apply elsewhere. We work with partners in Bangladesh and internationally, and most applied projects handle Bangla and English together.
Yes. That is a design goal, not an afterthought. Our models are built to run on-premise or in a private cloud region, on modest hardware, so sensitive data never leaves your control. We support AWS, Azure, GCP and on-premise GPU servers, and can work with self-hosted open-weight models as well as our own.
We design for data minimisation from the start: private or on-premise deployment, no training on your data by third parties, encryption in transit and at rest, role-based access, audit logging and, where required, data residency inside your country or cloud region. Data-sharing for research is governed by a written agreement.
Usually with a conversation about a hard problem and the data behind it. From there we agree a problem statement and evaluation plan, run a time-boxed pilot or proof of concept, and decide together whether and how to scale. Longer joint research projects and multi-year programs follow the same pattern.
A pilot or proof of concept typically takes 6 to 12 weeks. Moving a successful pilot into a production system that partners operate themselves usually takes a further 4 to 9 months in phases. We give a concrete timeline after the framing stage.
Partners own the systems, fine-tuned models, documentation and deliverables created for their project, and everything lives in repositories and infrastructure they control. The lab's own models such as Sangbad-SLM and its research assets remain ours; we agree the terms for using them up front.
Yes. We offer monitoring, model updates, evaluation in production and advisory support for systems that have left the pilot stage, and we train partner teams so they can operate and extend the work themselves.
Still have questions?
Tell us about your problem or dataset and the lab will reply within one business day.