May 2025 – Jul 2025 · 2 mos · IBM Research Bengaluru, Karnataka, India · On-site

• Developed a modular framework for automated Bash script evaluation for system level task and refinement that infers implicit preconditions, synthesizes test environments, and uses LLM-driven feedback loops to improve script correctness without requiring reference solutions
• Achieved significant performance improvements in script reliability with 5% accuracy gain on NL2Bash-EABench (83% vs 78%) and 3.4% improvement on Intercode-Corrections (56% vs 52.6%) compared to baseline methods, demonstrating superior evaluation capabilities over existing frameworks like execution-based (AgentCoder) and reference-based (CodeBERT, ICE-Score) methods.
• Designed and implemented execution-driven testing methodology that automatically generates targeted test cases and sandboxed environments, enabling practical deployment in production scenarios where gold-standard references are unavailable
• First Authored a paper on this work and submitted to AAAI 2026
• Filed for Patent for the same work