Tejas Ravishankar

Developing a low-code, drag drop platform to construct and automate complex agentic workflows.

Built a custom multi-agent state machine with human in the loop, step-based debugging and logging.

Developed 10+ software integrations, architected efficient React components to reduce client-side latency by 30% and created a real-time UI with websocket integration.

Architected a cloud platform provisioning secure Firecracker microVMs to users executing arbitrary code.

Developed a novel framework leveraging pretrained image transformers for video understanding tasks.

Utilized the NCut algorithm and Optical Flow to segregate dynamic foreground tokens from static backgrounds.

Treated video sequences as weighted graphs for robust temporal dynamics understanding, leveraging a graph transformer to maintain object continuity across frames.

Validated the framework using HMDB51 and AViD datasets, achieving ~9% lower accuracy than state-of-the-art while having just 1% of the trainable parameters.

Developed a web application using Angular, ASP.NET, and MongoDB, which was integrated with various APIs to enable users to stream and favorite videos.

Demonstrated proficiency in collaborative teamwork, employing version control, meticulous documentation, and comprehensive testing practices to ensure the successful development and maintenance of the project.

Enhanced low-light environments in real time video feeds using Cycle GANs and AutoEncoder-based models.

Experimented with image encoding in a memory buffer to reduce video frame transfer time over the network.

Prototype developed was under 1 megabyte and inference was in the order of 0.001 seconds per frame.

Adhered to ML Ops principles for efficient model deployment, monitoring, and continuous improvement with a focus on scalability, reliability, and reproducibility.

Designed a computationally efficient methodology for segmentation of narrow river streams in satellite imagery.

Utilized a combination of image processing methods - thresholding and gamma correction - and an encoder-decoder architecture constructed using depthwise separable convolutions.

Achieved higher accuracies than state-of-the-art models, whilst being 1/3rd the size.