Kaushik Koneripalli

Architected a bilingual Arabic/English LangGraph-based judicial reasoning multi-agent workflow and built a RAG pipeline across 1,000 legal cases and 10k+ evidence documents. Reduced end-to-end runtime from 10 minutes to 8 minutes for 10 concurrent users and achieved 84% overall accuracy using LLM-as-judge evaluation.

Developed a GNN-based reservoir simulator (SimGraph) for dynamic reservoir prediction and history matching, delivering up to 75× speedup over traditional simulation workflows. Built agentic and vision models including hierarchical patent intelligence RAG over 100K+ documents and a 3D U-Net for CT-based residual gas saturation prediction (0.88 IoU).

Built VAE-based generative models for pressure-vessel design under physical constraints and developed aircraft design data generation/evaluation pipelines using AWS and CI tooling. Implemented multimodal trojan detection baselines and an out-of-context satellite object detection pipeline, with work accepted at ICCV 2023 and evaluated using mAP and Mahalanobis-distance scoring.

Developed an asynchronous parallel Bayesian optimization framework with Gaussian Processes and Expected Improvement for set-based material discovery. Built ML systems for scientific diagram understanding and industrial computer vision, including deep learning pipelines for rail anomaly localization, defect detection, and 3D point cloud upsampling/object detection using modern detection frameworks.

Completed a Master of Science in Electrical Engineering from Arizona State University from 2017 to 2019.

Earned a Bachelor of Engineering in Telecommunication Engineering from PES Institute of Technology between 2013 and 2017.