Formal Methods & AI for Software Verification
I combine formal methods with machine learning and LLMs — neuro-symbolic SMT solving over bit-vector logic, CBMC-based bounded model checking and MC/DC coverage analysis of safety-critical embedded C/C++, and LLM fine-tuning for coverage-driven test generation. I aim to apply this to formal verification and AI-assisted EDA.
Co-developed a solver that combines machine-learned guidance with symbolic SMT reasoning over bit-vectors, with Lucas Cordeiro's group at the University of Manchester. Entered in the International SMT Competition 2026.
An SE-attention 1D-ResNet that ranks simulation-based test cases from road-waypoint kinematics — APFD 0.62 vs. 0.50 for random ordering, with a 100% completion rate across 400 configurations.
Benchmarked 7 LLMs on 42 programs (coverage, mutation score, cost, latency, hallucination rate) and tested generalization on 65 GNU Coreutils programs. LoRA-fine-tuned StarCoder2-7B and fixed a severe post-fine-tuning drop with a chat-format-aligned training strategy. Built a coverage-guided feedback loop that finds uncovered branches after each round and re-targets them.
Tool Paper · 19th IEEE International Conference on Software Testing, Verification and Validation (ICST 2026), Daejeon, South Korea · IEEE Xplore
Full list on Google Scholar.
Software Testing & Verification · NIT Warangal
GIET University
Open to research collaborations in formal methods, AI for verification and AI-assisted EDA.