Varun Badrinath Krishna

Varun Badrinath Krishna is a senior principal AI solutions engineer at SambaNova Systems and the creator of AttackQA, a cybersecurity question-answering dataset. His work spans smart-grid security, enterprise machine learning, and the infrastructure needed to run specialized language models efficiently.

From critical infrastructure to enterprise AI

Krishna earned an engineering degree from the National University of Singapore in 2010, worked as a research engineer in Singapore, and completed a master’s degree in 2016 and a doctorate in electrical and computer engineering in 2018 at the University of Illinois Urbana-Champaign. His doctoral research applied data-driven methods to electricity theft, compromised metering data, and smart-grid cyber-resilience. He held internships at ABB Corporate Research, IBM Research, Cisco, and C3IoT and joined the Siebel Scholars Class of 2018.

After working at C3.ai, Krishna joined SambaNova, where his interests expanded into retrieval, inference infrastructure, and multi-model enterprise applications. He participated with other SambaNova engineers in a 2024 AI Engineer workshop demonstrating Llama inference and document question-answering through embeddings, semantic retrieval, and vector indexing.

• AttackQA and cybersecurity question-answering. Krishna’s independently authored research transformed MITRE ATT&CK material into 25,335 question-answer pairs with supporting rationales, using multiple models to generate examples, reject weak material, and evaluate answers for security-analysis workflows.

• Fine-tuned retrieval and generation. AttackQA treats embeddings and answer generation as complementary optimization targets. On its specific benchmark, tuned open models outperformed the proprietary retrieval-and-generation configurations evaluated in the study.

• Model bundling for agentic workloads. Krishna advocates keeping specialized models on shared infrastructure, allowing retrieval, reasoning, validation, and synthesis to draw on different models without separate dedicated machines. His approach combines tiered-memory inference with shared hardware utilization to reduce handoff latency, infrastructure demands, and operational complexity.