Themes

Devices powered by ambient energy — solar, RF, vibration — operate intermittently. Conventional ML pipelines assume reliable power; the real world doesn't cooperate.

My work re-architects the entire stack: training procedures that bake intermittency into the optimizer, schedulers that handle partial computations, accelerators that adapt their shape to available power, and communication co-design that minimizes what gets sent off the node at all.

When the model is small relative to the dataset, the bottleneck isn't compute — it's the I/O. Computational storage moves work into the storage device itself, so encrypted, compressed, redundant data never has to cross the host bus before it's useful.

I build FPGA-accelerated CSDs that handle near-data compression, encryption, and redundancy for continuous-learning edge servers and parallel query pipelines.

A solar-powered edge server that doesn't need a battery is an architecture problem first and a power problem second. The accelerator has to morph as power fluctuates; the training procedure has to keep something useful happening even when the panel goes quiet.

Closely related work on retrieval-augmented generation pipelines for edge devices explores the same question for inference.

Whether the constraint is a public-cloud SLO or a router's power envelope, inference serving is a multi-dimensional optimization problem.

My contributions span ensemble-based serving in the cloud (Cocktail), serverless container provisioning for dynamic DAGs (Kraken), transfer-learning-based perf prediction (MLPP), and expert prediction for sparse mixture-of-experts inference (Prophet).

Domain-specific accelerators for AR/VR, point clouds, and ReRAM crossbars are where architecture meets perception. Spatio-temporal compute reuse for 360° VR streaming, octree-based point cloud compression at the edge, and analog-activation ReRAM neural network co-design are the venues where I've made the case that co-design beats general-purpose silicon for these workloads.