Roman Sinkus

• Architected and shipped a voice-to-voice AI sales-training platform now serving 50+ paying users and $5,000 in revenue, owning architecture, infrastructure, and customer-facing iteration end to end.
• Designed the retrieval-augmented LLM pipeline over 14+ hours of transcribed course content, letting users practice pitching, closing, objection handling, and mentality against a real-time conversational agent.
• Built the full stack: Next.js/React on AWS Amplify and a FastHTML backend on AWS App Runner via ECR with CI/CD and CloudWatch monitoring, integrating Ultravox for real-time voice, OpenAI for language modeling, and Supabase/Postgres for auth and data.

Next.js, React, FastHTML, AWS, Supabase, OpenAI, Ultravox

• Designed and developed a full-stack application to streamline midterm scheduling for the biochemistry and computer science departments.
• Implemented a force-directed graph algorithm to balance instructor preferences against scheduling constraints and reduce exam conflicts.
• Built an interactive interface with React, Node.js, and Tailwind CSS for real-time collaboration among instructors, improving workload distribution for students.

React, Node.js, Tailwind CSS

• Developed an integrated robotics simulation environment using ROS 2, Gazebo, and MoveIt 2, enabling detection and avoidance of 100% of robot-robot and robot-object collisions.
• Automated system setup with Bash scripts driven by user input, reducing configuration time by more than 10x.
• Collaborated with 6 faculty members to refine requirements and iteratively design the robotics software architecture.

ROS 2, Gazebo, MoveIt 2, Bash

• Hardened API security by migrating sensitive GET endpoints to POST, keeping credentials and PII out of URLs, server logs, and browser history.
• Reorganized the regression testing plan by fixing and augmenting 10+ broken end-to-end Selenium tests, saving the development team 15+ hours each two-week sprint.
• Spearheaded the effort to achieve W3C WAI compliance for 100% of React components in a hospital screener application.

TypeScript, React, Node.js, Express, Selenium

• Decreased video decoding times by 10x by enabling ffmpeg HEVC hardware acceleration using CUDA.
• Created and managed Git submodules and Debian packages to integrate third-party technologies.
• Extended and maintained Evertz's internal media-processing API in C++, validating requests and responses with Boost unit test suites.

C++, CUDA, ffmpeg, Boost

• Led the computer vision subteam building an autonomous service robot for the RoboCup@Home competition.
• Developed a ROS service using the face-recognition library to store facial data and identify individuals accurately.

ROS, Python, face-recognition

• Built a medical logbook transcription app with UBC medical students, digitizing handwritten notes using TypeScript, React, and open-source LLMs.
• Shipped a restaurant recommendation app in React Native and an epilepsy patient tracking app in Flutter across earlier project cycles.

TypeScript, React, React Native, Flutter, Dart

• Worked with a postdoctoral fellow on a predictive model identifying Parkinson's patients from spiral drawings.
• Analyzed spiral drawings in Python, using approximate entropy to quantify disorder in the drawings.
• Presented at UBC's Multidisciplinary Undergraduate Research Conference and received the REX Innovation and Technology Award.

Python, approximate entropy

• Worked with an assistant professor exploring research questions in chatbot development.
• Analyzed and visualized chatbot data with NLTK and spaCy, laying the foundation for a research paper.

Python, NLTK, spaCy, Jupyter