Simon Gilbert Email & Phone Number

New York, Ny, Us

New York, Ny, Us

San Francisco, California, Us

Singapore, Singapore, Sg

New York, Ny, Us

● Spearheaded machine learning and software development for Digital Coworkers. -This end-to-end document classifier and data extractor is currently in production and processes ~3000 tasks/day, doing the work of 12 people, and saving ~500,000 items/yr for future ML projects. -The classifier achieves 99.6% accuracy across 15 labels and extracts existing claim information with 97% accuracy. -Demonstrated the desired functionality with a POC, acquired the customer, and created a production ready implementation deployed using Azure services over the course of 6 months. -Set up alerting, metrics, and automated retraining in a CI/CD pipeline. ● Contributed to machine learning/design efforts for latest company product: Roots Flow - a customizable framework for AI-based insurance claims indexing. -Product reads files, extracts information based on a custom config, and trains ML models with minimal developer effort. -Roots Flow uses a DAG-based data orchestration workflow which automatically gathers and stores data to train and run ML models. -Data is available across customers and will be used to create a universal claims information extractor.● Created error correction tool to predict bounding boxes missed by OCR software with 99% accuracy. Built several models, found best performing, and integrated into Roots software.● Coded AWS pipeline to asynchronously extract text from documents.● Researched state of the art semantic matching models and developed POC using word embeddings, BERT, and other transformer-based architectures. Achieved 87% accuracy across 12 classes.● Developed error-catching random forest model which predicts when a user incorrectly inputs data into a form with 97% accuracy. Analyzed most important features to create additional inputs to improve accuracy.

Boston, Ma, Us

• Designed and built end-to-end machine learning system for clinicians looking to provide idiopathic pulmonary fibrosis patients with an accurate prognosis: cleaned data for model training, analyzed dataset for biases, surveyed segmentation models, and built GUI.• Used transfer learning to adapt U-net for de-noising CT scans of COVID patients’ lungs in order to enhance scan clarity and decrease patient x-ray exposure.• Created image captioning models with novel encoder/decoder pairs and evaluated performance using BLEU. o Implemented transformers with multi-head attention from scratch to improve caption generation performance by 11% over traditional model. The team tested a total of 6 architectures with multi-head attention being the best performing!

• Coded deep convolutional neural network to create 3D from 2D hologram using a scattering model as a prior.• Trained denoising models on multiple >10¹¹B datasets and created figures analyzing performance.• Evaluated network success using self-coded Dice metric and tuned network parameters to optimize results.• Research paper pending on particle localization using deep priors trained on simulated data.

Brooklyn, New York, Us

Taught three sections of modeling-physics, one section of AP Physics 2, and one section of Advanced Topics in PhysicsHead-coached Science Olympiad TeamAdvised an advisory group of nine 9th graders and co-advised a larger homeroom of eighteen 9th graders

Teach 3 sections of a conceptual physics classTeach an introductory VEX Robotics classTeach an advanced VEX Robotics classSupervise the robotics club and coach robotics competitions for both middle and high schoolOrganize and advise high school Science OlympiadCo-advised an advisory group consisting of 14 10-12th gradersAssist with upper school jazz band and orchestraOversee Super Smash Brothers Club

Master Of Science - Ms, Computer Engineering

B.A., Physics

Education record