Work

Post-training Granite LLMs at scale: Training instruct/reasoning LLMs with RL across recipes, algorithms, stacks and domains to optimize reasoning & agentic performance.
Inducing efficient reasoning capabilities in LLMs: Implemented techniques in the RL training pipeline for verbosity control and token efficiency in large models.
Optimizing and stabilizing training stacks for RL: Developing features and fixes on top of open-source RL training stacks (open-instruct/skyrl/nemo-rl) to make them stable and scale up large language model training while maximizing train-time throughput.

Built an agentic service for pipeline failure detection and resolution: Developed an intelligent ML-based service for automated failure detection and resolution, utilizing RAG Agents and vector search over a dynamically constructed, continually growing knowledge base.
Devised a custom controllable cloud-native CI/CD pipeline framework: Designed microservices to replace legacy Jenkins infra for improved resource allocation during job allocation.
Optimized execution timings of long-running regression (LRG) tests: Introduced parallelization of LRG tests in monolith applications considering unit test inter-dependencies.

Transferable Knowledge Systems | Prof. Abir Das, Prof. Saptarshi Ghosh (CSE, IIT Kharagpur)

Developed an on-the-fly prompting technique with language attribute similarity computation, addressing catastrophic forgetting in class-incremental continual learning while also cutting parameter overhead by 50%. (Paper @CVPR '24)
Designed a fully source-free, actively controlled student-teacher based knowledge refinement mechanism to perform test-time adaptation with continually changing target domain distributions.