Turkish defence firm Havelsan introduces indigenous AI-based assistant

Turkish defence firm Havelsan developed the country’s artificial intelligence (AI)-based virtual assistant specialising in natural language processing, scheduled to be made public on February 1.

The AI-based assistant runs on the “indigenous GPT model,” or Generative Pre-Trained Transformer, a ground-breaking natural language processing model. It is trained on diverse data to deliver a wide range of text-based tasks, or depending on the version, it can process and understand multimedia.

Osman Kavaf, the product engineer for Information and Communication Technologies at Havelsan, told Anadolu Agency that the firm is conducting different projects on AI, developing specific algorithms on text and multimedia.

Havelsan mostly proceeded on a project basis and started working on GPT models a year ago in anticipation of US-based OpenAI’s model, ChatGPT, Kavaf said.

He said the first version of the GPT model is ready, and the name chosen for the model, MAIN, is an abbreviation for Multifunctional Artificial Intelligence Network.

The assistant's feature of natural language processing, or NLP, refers to a computer’s ability to use and understand spoken and or written language akin to a human.

“We, as Havelsan, agreed to take steps to ensure data security, especially considering the needs of public and military institutions, and we followed through with the project, as the first version of our indigenous GPT model is planned to be launched on February 1,” he added.

Features of the indigenous model

The GPT model developed has the basic features of all models available for access, one of which is summarisation, according to Kavaf.

He assured that currently, the Havelsan-developed GPT can also generate computer code, allowing users to input different parameters to make iterations on the output.

He said the project was launched to ensure the data security of public and military institutions, adding that they want to develop a mechanism to provide public service in future versions.

Data privacy

Regarding how the GPT model is accessed, Kavaf said it needs a proprietary hardware kit, a type of particular device like a “closed box,” to prevent data leaks.

Acknowledging that the way Havelsan employs the product is unlike any other, Kavaf said the company wants to use the model in all of its products. Although outside the privacy goals, it may integrate cloud systems for remote access to the AI model in the future, he added.

“Cloud integration does not stand out as a solution when handling sensitive and highly confidential data,” according to the engineer.

He emphasised that Havelsan is working with a project group of 20, mainly academics, and said the number of people involved is expected to gradually increase as the company plans to work with engineering students from universities with which they have agreements.

Kavaf stated that the company is planning a big event for AI for launch day on February 1 at Havelsan, as it will then be showcasing the GPT model.

“We will start seeing AI in many different areas of our lives as the world becomes more aware of its potential. This first release of our model will be more text-focused, and we will add image and audio processing in future stages,” he added.