Suman Mandava

• Ported the Qwen2.5-VL Vision Encoder to Kornia in native PyTorch (PR #3409), implementing Conv3D patch merging, rotary embeddings, and attention modules — eliminating HuggingFace dependencies for production deployments. • Proposed and led VLM/VLA model support initiative, opening Kornia's multimodal architecture roadmap for future contributors. • Designed modular, type-safe components and resolved CI failures to meet production-quality merge standards.

• Built an end-to-end affective state recognition pipeline on DAiSEE using multi-head Transformer encoders over Action Unit (AU) and Valence–Arousal (VA) features, achieving 78.08% test accuracy via late fusion. • Fine-tuned Qwen2.5-1.7B with LoRA using AU-based and rule-based prompts; trained per-label adapters with constrained decoding, improving engagement classification accuracy from 0.24% to 42.98% (+42.74pp). • Deployed an interactive inference demo on Hugging Face Spaces for real-time AU-driven emotion prediction visualization.

• Engineered real-time distributed HMI applications for metro control systems (BMRCL, DMRC) using ZeroMQ, protobuf, and WebSockets — reducing system latency 30% under high-frequency concurrent data loads. • Built operator dashboards in Node.js and Handlebars.js, translating hardware safety requirements into production UI features across two sprint cycles. • Partnered with cross-functional hardware and software teams to align HMI specs with operational standards, reducing integration rework across delivery phases.

• Built a full-stack Online Video Platform (OVP) using React, Node.js, and MongoDB. • Implemented role-based access for admin and client users, improving usability and access control. • Developed JWT-based authentication, video streaming workflows, and user analytics features to enhance engagement and platform reliability. • Followed modern development practices including RESTful API design, version control, and debugging in a collaborative environment.