Pixels to Layers
An architecture that turns generated infographics into editable, layered assets. Oral at the AI4CC Workshop, CVPR 2026 (non-archival) and accepted to CVPR 2026 Demos.
Research prototypes and applied systems across safety, vision, language and optimization.
An architecture that turns generated infographics into editable, layered assets. Oral at the AI4CC Workshop, CVPR 2026 (non-archival) and accepted to CVPR 2026 Demos.
Learned Editable Graphic Outputs — an expert-guided orchestration framework for adaptive infographic decomposition into editable graphic outputs.
A model-in-the-rear guardrail framework for safeguarding LLMs at low cost and scale. Won 1st place (solo) at the CNS Hackathon.
Context-aware memory modules with efficient lightweight vision transformers for image segmentation. Oral Accept at MIUA 2026 (to appear); 2nd prize, ACM Hackathon.
Large-scale BERTopic analysis of the LMSYS-Chat-1M dataset to understand human preferences across competing LLMs.
A YOLOv8 detector for cattle disease reaching 96% accuracy, with an oriented-bounding-box pipeline for industrial grain quality analysis.
Fine-tuned Google mT5 for English→Kannada transliteration (74% accuracy) for a Karnataka Government documentation project.
Point-cloud analysis and computer-vision for spatial mapping in an autonomous spacecraft-docking project at ISRO LEOS.
A co-designed, landscape-aware optimization system bridging serial intelligence and GPU parallelism (EvoApplications 2026).
A crossover-aware ML scheduler benchmarked across heterogeneous compute, balancing throughput and fairness.
Inventors: A. Bhandarkar, V. N. Balasubramanian, Gongjie Qi, Bala Kumaravel
A patent arising from research at Microsoft Research, currently in the process of grant.
Application No. 202541017132 A
A novel algorithm for efficient probabilistic string search with applications in data retrieval.
Application No. 202541017133 A
A sustainable energy solution integrating microbial and hydrogen fuel-cell technologies.