Projects

This section showcases a mix of personal and professional projects, including those from my time working with enterprise customers at AWS ML Solutions Lab (now Generative AI Innovation Center). Projects listed reflect my contributions both as a direct contributor and as a lead or manager.

All AWS-related projects included here are public knowledge. The inclusion of AWS or customer names does not imply any endorsement. The descriptions provided are my own, based on my own role and experience.

RowMaker

RowMaker is a mobile app that lets you extract rows of data from images and voice transcriptions directly into Google Sheets, making data entry portable.

Under the hood, RowMaker uses VLMs for multi-modal inference, processing both images and voice notes to extract structured data based on user-defined fields.

This application is built on VLMs, utilizing:

Multi-modal inference
Structured data extraction
Function calling and API integration
Zero-shot learning for new data types

RowMaker is the latest SIMPLtech app, designed to harness Gen-AI for personal productivity.

Learn more about RowMaker and try it for yourself at https://www.simpltech.ai

NoteTech

NoteTech is a mobile app that lets you build personal automations and tools ("tech") in under a minute by simply writing notes.

Under the hood, NoteTech creates autos with LLMs by generating directed acyclic graphs (DAGs) from modular serverless function nodes, using the user's note as input.

This application is built on the modern LLM stack, leveraging:

Zero-shot learning
Structured output generation
RAG
Agentic workflows.

NoteTech is the first app from SIMPLtech, a company founded to develop practical consumer apps with LLMs.

Learn more about NoteTech and try it for yourself at https://www.simpltech.ai/.

Blockitect

Based off NoteTech, Blockitect is an open-source library for generating executable DAGs (Directed Acyclic Graphs) from modular "building block" functions and NL prompts.

Users describe automations, and an LLM assembles them into workflows using predefined functions.

How it works:

Developers create modular "building blocks" such as data operations, API calls, or transformations.
These functions are initialized within Blockitect.
Users describe automations and the LLM generates an executable DAG that represents the automation flow.

Once configured, users can build self-service automations simply by describing their workflows in plain language, with Blockitect handling the creation of the executable structure.

NFL Player Health & Safety (AWS)

Helped close NFL-AWS partnership on Player Health & Safety by establishing innovative capacity
Developed a dense pose estimation model for approximating the kinematics of player-to-player collisions in American football.
Implemented Mask-RCNN to map pixel values to 3D mesh surface coordinates.

NFL-AWS Player Health & Safety Partnership

Tyson Industrial Automation (AWS)

Lead team to develop a computer vision solution for Tyson to count chicken trays during packing
Fine-tuned an SSD with ResNet 50 backbone using an end-to-end labeling and training pipeline on SageMaker
Deployed the model to an edge device for real-time inference offline

AWS Blog

Gilead Sciences Enterprise Search

Lead team to rapidly develop a prototype for enterprise search across Gilead's pharmaceutical development and manufacturing (PDM) business unit using Amazon Kendra
Subsequent deployments reduced information retrieval time by 50% for Gilead's PDM team.

AWS Case Study

Zero-shot virtual product placement (AWS)

Advised team in developing a novel framework for zero shot virtual product placement
Pipeline for identifying ad-insertion locations in cooking shows, placing ads along insertion planes, and rendering across frames w/ lighting
Zero-shot pipeline entirely on video data w/o camera parameters

AWS Paper

MLL Leukemia Diagnostics (AWS)

Lead team in collaborating with the Munich Leukemia Lab to explore how ML can be applied to next-gen sequencing data to accelerate hematological diagnostics
Collaborated w/ MLL bioinformaticians to engineer biomarker features based on the literature
Trained a LightGBM with Bayesian HPO to classify 30 subtypes of leukemia from 800 features

AWS Blog

Cyclical Learning Rate for Keras (Github)

Implemented a Keras callback for training with cyclical learning rate policies, as detailed in Leslie Smith's paper Cyclical Learning Rates for Training Neural Networks
A cyclical learning rate is a policy of learning rate adjustment that increases the learning rate off a base value in a cyclical nature.

Github

PGA Tour Putting Tracker (AWS)

Lead team in developing a multi-stage CV pipeline for tracking golf ball putts in 3D coordinates from video.
Combined fine-tuned YOLO V7 and off-the-shelf models to localize the putter and ball in stages, before using pre-trained single-object tracking for the ball
Transformed ball pixel coordinates to 3d coordinates using pre-computed transformation matrices via point cloud

AWS Blog

Brain MRI Segmentation @ Edge (AWS)

Developed prototype for brain MRI segmentation at the edge using a Raspberry Pi
Trained semantic segmentation models (U-Net, ENet) from scratch to annotate brain tissue types
Implemented cloud-to-edge pipeline for training models remotely and deploying to an offline environment

AWS Blog 1, AWS Blog 2