Curriculum Vitae — Sreevatsank Kadaveru

Communication Theory and Systems | PI: Dinesh Bharadia

Communication Theory and Systems | ECE Department

Technology Management Program

Electrical Engineering

Developed a rigorous framework to study LLMs for RF+Vision based multimodal sensing tasks, to identify and isolate failures to gaps in domain knowledge, domain reasoning or domain-specific data interpretation, and prescribe targeted fine-tuning remedies.

Blind Detection and Tracking of Low Probability of Intercept Signals.

• Developed a novel statistical approach for wideband spectrum sensing, improving state-of-the-art prediction and enabling multi-receiver capability via an optimal cooperative scheduling algorithm. [Poster]

Developed highly automated CI/CD Kubernetes workflows improving reliability and efficiency by 90%.

Improving Reliability of MIMO Mission Critical Systems using LDPC in 5G.

• Demonstrated a novel latency-reliability trade-off in LDPC code selection for video traffic

• PAIQoO: Quantifying the network requirements of Physical AI applications
  • Characterizing how sensor fidelity and end-to-end latency trade off in robotic task performance
• mmsRAN: Radio Access Networks for Physical AI [Paper]
  • Building an Uplink Scheduler for Multimodal sensor flows that need to be low latency, synchronized and in lock-step order.
• P2SIF: Plug-n-Play Sensor Integration Framework [Proposal]
  • Developed a decoupled architecture for new modality integration into existing LLMs without needing to modify LLM weights

• PHYNet: A gold-standard PHY layer labelled dataset
  • Establishing a standard for evaluating PHY layer algorithms by building a multi-purpose labelled PHY dataset
  • Studied the problem of PHY spectrum ignorance in today’s world
• CamoPHYlage: Hiding signals in plain sight
  • Developed adversarial mimicking techniques to evade all PHY layer detectors, enabling surreptitious communication in plain sight

• rfYOLO: Phase modality for better detection
  • Developed phase transform to improve PHY layer characterisation of signals using YOLO networks
• Nomad: Architected a high throughput spectrum sensing system with a line rate of 40Gbit/s [Link]
  • Developed containerised micro-services with well-defined high throughput shared memory interface
  • Achieved major throughput gains with the use of copy-efficient shared memory and a Redis-based in-memory database.

• Curriculum FL: Studied effect of curriculum learning in a federated context. We showed that curriculum can significantly improve convergence time of training. [Paper]

• Audio-StyleCLIP: Extended capabilities of StyleCLIP and AudioCLIP models to create a model that can perform style transfer onto an image using an audio styling input. [Report]
• Convexifying Neural Network: Implemented the state-of-the-art technique for modelling a ReLU neural network as a convex optimisation problem, the solution of which is a globally optimal neural network. [Report]
• Training Robust NN: Implemented Lipschitz constant based regularization for training robust Neural Networks and compared performance against L₂ and other baseline techniques. [Report]
• Lattice Basis Reduction Performance Study: We evaluate the basis reduction algorithms, such as LLL, BKZ, Slide, Self-dual BKZ, to study accuracy vs performance trade-off on COTS hardware. We also study scaling of these algorithms with problem size. [Report]