Srihari Unnikrishnan

Machine Learning Researcher · Deep Learning · NLP · Systems for AI

About

I am a machine learning researcher working on the theoretical and empirical foundations of deep learning systems, with a particular focus on transformer architectures, interpretability, and reliable deployment of large language models. My work spans mechanistic analysis of attention, agentic systems, and data-centric AI, with applications in cloud systems and structured data.

Research Interests

• Transformer architectures and attention mechanisms
• Interpretability and mechanistic analysis of LLMs
• Agentic systems and tool-augmented language models
• Numerical stability, optimization, and efficient training
• Robustness, hallucination detection, and model evaluation

Experience

• AI Researcher, Microsoft Research India
• AI Research Scientist, Sony Research India
• Data Scientist, PricewaterhouseCoopers India
• Freelance Research Engineer (ML systems & applied AI)

Selected Research & Projects

• GPT From Scratch — A from-first-principles implementation of a GPT-style transformer to study attention, causal masking, and training dynamics at a mechanistic level. Emphasizes architectural constraints, numerical stability, and GPU efficiency.
  Project Page · Code
• TokenBurn — A library for LLM confidence estimation and hallucination detection using log-probabilities, perplexity, entropy, and KL divergence. Evaluated on the HaluEval benchmark.
  Code
• TabGuard — An agentic framework for adaptive tabular anomaly detection via dynamic validator selection and programmatic execution. Developed at Microsoft Research; outperforms existing benchmarks.
• FixItFlow — Automated troubleshooting guide generation from cloud incident logs. Extracts diagnostic patterns from historical engineer actions and synthesizes validated remediation guides, achieving significant reductions in mitigation time.
• Liquidity-Aware Bond Yield Prediction — Hybrid CausalGAN and reinforcement learning approach for synthetic bond yield generation using macroeconomic variables.
  arXiv

Education

B.Tech in Computer Science and Engineering (Artificial Intelligence & Machine Learning)
Vellore Institute of Technology, Chennai