Toby Savage Email & Phone Number

Carlsbad, California, United States

• Built an advanced LLM system that generates descriptive captions from orthopedic sensor data, providing real-time insights to healthcare professionals in physical therapy and streamlining patient progress assessment for improved recovery support.• Fine-tuned GPT-2 on over 4,500 images through Databricks, leveraging CUDA-enabled GPU training and an MLP to map image embeddings into the LLM's text space.• Currently working to boost output quality and raise evaluation scores, aiming for higher specificity and accuracy with current METEOR and SPICE scores around 0.5.• Evaluating performance using BLEU, METEOR, ROUGE, and SPICE metrics to fine-tune the system and better align outputs with reference captions.

Carlsbad, California, United States

• Led and developed a customer service agent using OpenAI’s GPT models and proprietary machine learning models to automatically answer patient questions related to Canary’s smart implant for Total Knee Arthroplasty.• Prompt engineered 'gpt-3.5-turbo' model to shape behavior, utilizing context and instructional prompts, and few-shot techniques. • Trained a Naive Bayes classification model on Frequently Asked Questions data to predict the category of a user’s question using OpenAI's embeddings model, 'text-embedding-ada-002'. Categorization is used to retrieve prominent information from a separate database to guide the GPT model to answer the user’s question, drastically minimizing hallucinations.• Compared performance of current implementation of customer service agent with Pinecone Vector Database.

Carlsbad, California, United States

• Calculated key metrics such as Patient Compliance Rate and Implant Failure Rate using SQL and Python, identifying missing data causes to aid executives in diagnosing compliance or implant functionality issues.• Developed and deployed a Microsoft Azure-based API for kinematic 3D animations for internal dashboard,facilitating detailed performance studies of knee replacements post-surgery.• Enhanced critical performance dashboard to compute live key metrics and readily express data completeness and quality.• Augmented internal dashboard to improve corporate operational efficiency.

Boulder, Colorado, United States

• Led a team of 18 software engineers to design and construct software architecture, networking, data manipulation, and navigation systems for Tunnel Boring Machine (TBM), following an agile development methodology.• Pioneered software team through The Boring Company's Final Design Package (FDP) round; accepted as one of 12 teams out of 400 worldwide to compete in the Not-A-Boring Competition for Summer 2021.• Produced Kalman Filter for TBM to predict underground positioning based on IMU sensors and actuator data.• Drove the implementation of a server-client network connection between ground station computer and an Arduino Portenta H7 board, sending UDP packets through an ethernet cable.• Designed algorithms for actuator functionality and networking for TBM.• Organizing team discussions and action items while following an agile development methodology.

Boulder, Colorado, United States

• Built a collaborative text editor application, called Bingo Bongo, using TypeScript, allowing multiple users to create accounts, update passwords, and edit a single document simultaneously.• Implemented an Object-Relational Mapping tool called Zapatos with collaborative editor to enable a service to query user and document data from Postgres database.• Enabled client-server networking communication within collaborative editor through HTTP requests.• Gained experience working with Javascript and Ruby on an application called Helios.• Implemented a subscription for SHA values for Helios that checks if the client has the most up-to-date version. • Refactored Helios's Higher-Order Component, replacing repetitive GraphQL queries with fragments for application components.

Carlsbad, California, United States

• Debugged and ran unit tests on world’s first smart knee implant algorithm using IMU sensor telemetry data, including 3D high-frequency gyroscope and accelerometer data.• Generated requirements based verification test document results for FDA submission. Company received an FDA De Novo approval grant one year later.• Contributed to the design of the company's database architecture for patient and implant data for better organization.• Self-taught and debugged MatLab code through analyzing custom algorithms for implant calculations.• Tested networking between operating room applications, base station hardware, and implants.• Analyzed networking logs between base stations and implants. • Aided select surgeons performing implant trial procedure as a Circulating Nurse; collected data regarding the performance of implants during the company’s pivotal cadaver study.

Tinton Falls, New Jersey, United States

• Optimized statistical analysis on plant species abundance, including Total Percent Cover, Floristic Quality Assessment Index, and Plant Stewardship Index, by developing an R script that reduced computation time by over 75%, significantly enhancing data processing speed compared to manual calculations.• Assisted in data collection and statistical analysis on multiple long term forest health monitoring projects: Deer Exclosure Study, Hartshorne Woods Orchid Study.• Utilized Microsoft Access database to collect data on Deer Exclosure Study and Hartshorne Woods Orchid Study, Wildlife Reports.• Created Trail maps for local parks using Geographic Information Systems (GIS).

Master Of Science - Ms, Computer Science, Machine Learning

Bachelor Of Science - Bs, Computer Science

Activities and Societies: CU Hyperloop• Cum Laude • Engineering Merit Scholarship • Dean’s List