Intel

• Led R&D into a novel GenAI feature that improves hardware design flows by leveraging LLM design generation.
• Built a data processing and augmentation pipeline to create a hardware system dataset with millions of tokens.
• Outperformed SOTA prompt-engineered foundation models by 300% on test set by leveraging performance efficient fine tuning (QLoRA) and RAG to fine-tune an open source LLM with Transformers library on Azure ML Studio.
• Designed a custom language for LLM outputs, increasing total inference speed by 100% from baseline.
• Developed and integrated C++ and Tcl code to support the migration to a modern database for Quartus tools, using Object Oriented Programming patterns for improved maintainability.
• Architected a testing framework for a user-facing debug tool and used it to validate the tool’s reliability on thousands of inputs from customer designs.
• Continuously resolved customer requests and bugs by implementing software features and fixes, including creating a Quartus Python module for merging customer debug files.
• Designed a custom language for LLM outputs, increasing total inference speed by 100% from baseline.
• Tools: C++, Python, Transformers, AzureML, Bash, Tcl

The Matter Lab

• Developed a pipeline to process 40,000 datapoints with xtb, Open Babel, and PyTorch, building a molecule dataset. dispenser robots; improved an optimization algorithm that reduced the cycle-time of the automation process by 25%.
• Trained and tuned graph convolutional neural networks to predict molecule solubility with Pytorch Geometric while experimenting with various architectures and physical chemistry features
• Demonstrated that a electronic surface mesh approach is valid on solubility task, beating some architectures such as XGBoost and GC by 10% in RMSE loss on ESOL solubility task
• Tools: Python, PyTorch, Bash, Pyvista, Openbabel, xtb

Rocscience Inc.

• Developed a program in C++ that fixes mesh holes by transforming meshes into voxel grids with OpenVDB.
• Improved triangle element feature query speed in C# by more than 10x by redesigning algorithms to leverage Bounding Volume Hierarchies.
• Collaborated on and presented a cross-platform model viewer in Typescript using Vue.js and Electron.Bounding Volume Hierarchies.
• Developed an augmented reality interactive 3D geotechnical model viewer app using Unreal Engine 4 and deployed it on the Microsoft HoloLens.
• Tools:C++, C#, Javascript/Typescript, Vue.js, Node.js

Languages and Databases

Python

C++

C

Javascript

C#

Bash

Libraries

Pytorch

Transformers

NumPy

Pandas

JAX

Frameworks

React

Flask

Vue.js

Node.js

Technologies

Git

Perforce

Linux

Docker

Azure ML