This project showcases the functional verification and live demonstration of a custom silicon chip implementing a Recurrent Spiking Neural Network (RSNN)—a bio-inspired architecture designed entirely through natural language prompts to a Large Language Model (LLM), specifically ChatGPT-4.
The RSNN was synthesized from Verilog code generated by ChatGPT, validated on tasks like XOR, IRIS classification, and MNIST, and ultimately fabricated through the open-source TinyTapeout ASIC shuttle using SkyWater’s 130nm process.
To verify the fabricated chip, our team developed a complete hardware-in-the-loop test framework using microcotb, a Python-based cocotb-like environment adapted for embedded systems. We also created custom MicroPython firmware for the Raspberry Pi Pico to enable direct communication with the chip for streaming neuron parameters, injecting spike inputs, load test scripts and reading output spike trains in real time.

SurgiVision explores the use of lightweight artificial intelligence to interpret surgical scenes in real time. The project aims to develop a system that can identify key elements during minimally invasive procedures—such as surgical tools, actions, and target anatomy—using visual data. Designed with efficiency in mind, this approach is intended for potential deployment on resource-constrained platforms like embedded or edge devices. The project aims to explore how artificial intelligence can support surgical workflows in clinical or educational environments, such as in smart operating rooms, robotic assistance, or surgical training.

PbS colloidal quantum dots (PbS CQDs) are a promising material for next-generation solar cells and thus are a focus of research for renewable energy. The tunability of PbS CQD sizes allows for precise control of bandgap energy in the visible and near-infrared region, particularly useful for creating multi-junction solar cells. Current commercial solar cells use bulk material in a single-junction architecture, meaning only a certain range of wavelengths of light from the sun is absorbed.

Multi-junction solar cell architectures utilize multiple layers with different properties that absorb across a wider sunlight bandwidth, allowing for power conversion efficiencies that surpass the limits of a single-junction cell. With the goal of building a multi-junction architecture, we aim to optimize the synthesis of CQDs with bandgap energies of 1 eV and 1.6 eV. We also present a novel synthesis method for a single-population of ultrasmall PbS CQDs with an exciton range of 600-800 nm, useful for triple junction solar cells.

ScolAI is a machine learning pipeline designed to automate the measurement of the Cobb angle, a key metric for diagnosing scoliosis. Manual Cobb angle measurements are time-intensive and prone to variability between clinicians, which can affect treatment decisions. Our approach uses AP X-rays from scoliosis patients and implements two main strategies: hip segmentation as a proxy task and direct angle regression. The segmentation model is built on a U-Net architecture, while the regression task leverages ConvNeXT, Swin Transformer, and ResNet-50 to predict Cobb angles from spinal X-rays. Our models demonstrate low error rates and strong generalization, setting the stage for future deployment on labeled spinal datasets. ScolAI holds promise for streamlining clinical workflows and improving consistency in scoliosis care.