Jamnagar (Gujarat), India, November 21: Parshu Tactical Defence Pvt. Ltd., an Alwar-based deep-tech startup incubated at IIT Madras, has trained more than 56 Indian Navy officers in hands-on fine-tuning of large language models (LLMs) at INS Valsura, home to the Navy’s Centre of Excellence for Artificial Intelligence and Big Data Analytics.
According to the company, this was the first time the Indian Navy conducted LLM training that was entirely hands-on and practical rather than theory-based.
The three-day workshop, titled “Fine-Tuning of Large Language Models for Maritime Domain Awareness and Decision Support,” ran from November 16 to 18. By the end of the program, every officer had launched a model, trained it over multiple epochs and deployed it.
The emphasis was on practical experience rather than lectures. Working in small teams, officers completed the entire process themselves, including hosting models on GPUs, loading raw base models, cleaning raw data, fine-tuning models over successive epochs and building retrieval-augmented generation (RAG) pipelines that enable models to generate responses based on approved documents.
The training was led by Parshu Founder and CEO Priyanshu Joshi, along with AI specialists Shivanshu Sharma and Ravi Raj. INS Valsura honored all three with mementos.
“We did not want officers to watch us train a model; we wanted them to do it themselves,” Sharma said. “By the third day, every officer had taken a raw base model all the way to a deployed, working system.”
“Security was built in from the first step: Each team worked in its own isolated GPU workspace, using only unclassified and synthetic data,” Raj said. “The officers now have a method they can repeat on the Navy’s own air-gapped systems, where real operational data never leaves naval networks.”
Joshi is an AI specialist with more than 15 years of experience, including 13 years in Dubai. He holds a B.Tech. in Computer Science and an MBA from Hult International Business School.
He recently co-authored the United Service Institution of India (USI) occasional paper, “Unmanned Surface Vessels in Contemporary Maritime Conflict,” with retired Lieutenant General Karanbir Singh Brar, PVSM and AVSM. The paper examines how unmanned surface vessels (USVs) were used in conflicts between 2022 and 2026.
The training addresses a practical need. Models fine-tuned for naval terminology can help officers quickly analyze large volumes of reports, logs and signals. However, sensitive naval data cannot be shared with public AI platforms. The Navy therefore needs officers capable of adapting open-weight models for maritime applications and operating them on systems under its full control.
“Sovereign AI will be built by people who can train, test and deploy models themselves,” Joshi said. “We are grateful to INS Valsura and the Indian Navy for their trust.”
Joshi said the workshop is the first in a series of hands-on AI training programs that Parshu plans to conduct for the armed forces, including the Indian Navy and the Indian Army. The company also plans to offer the program to businesses seeking to train their employees to fine-tune and deploy AI models on their own infrastructure.
Incubated at IIT Madras Pravartak and the IIT Madras Incubation Cell, and recognized by the Department for Promotion of Industry and Internal Trade (DPIIT), Parshu Tactical Defence is developing Level 4 autonomous unmanned surface vessels for Indian waters and export markets.
The company describes its vessels as the world’s most advanced Level 4 USVs, featuring fully autonomous physical AI technology for maritime operations. They are designed to perform long-range surveillance, one-way attack and interception roles. Its technology demonstration vessel is currently under construction.
INS Valsura also shared updates on X, formerly Twitter, about the three-day hands-on LLM training workshop conducted in collaboration with Parshu Tactical Defence Pvt. Ltd. The post highlighted the participation of more than 56 Indian Navy officers in practical training on building, fine-tuning and deploying large language models for maritime domain awareness and decision support.
