Cenyue Yao

• Containerized Hunyuan and Trellis with Docker and resolved CUDA and PyTorch mismatches. Cut the first full environment setup from nearly a day to a consistent half day, a 50% reduction under a limited outbound network, and made the pipeline reproducible on the remote server.
• Standardized inputs and key parameters across 5 image-to-mesh pipelines, and moved all tests onto one shared template on the same scenes. Trimmed duplicate configuration steps across 5 tools and turned ad hoc comparisons into a single apples-to-apples workflow across 5 of 5 tools.
• Built 8 evaluation sets covering bright and low-light conditions and the 4 vehicle viewpoints: left front, left rear, right front, right rear, and completed 8 of 8 planned single-view runs with Colmap and in-house utilities for mesh and 3D Gaussian Splatting checks, achieving 100% coverage of the phase test grid.
• Ran 20 randomized parameter combinations in Trellis multi-image mode and documented that seed choice drives global layout while other knobs mostly affect shape and detail. This reduced unproductive reruns in later trials by about 30% through better upfront seeding and presets.

Serverless Apps on AWS Movie Search and Number to English

2025.05-2025.07

• Delivered two serverless mini apps powered by three AWS services. Shipped 2 REST endpoints and two demo pages, producing one deployment guide and one user note.
• Built the Movie Search Engine and ingested about 5,000 records into OpenSearch, writing one Python Lambda and configuring one API Gateway endpoint and one search page. Achieved about 1.0 s median end-to-end time.
• Validated search quality by designing 20 keyword cases. Reached about 90% title hit rate and completed 10 rounds of end-to-end stability checks with zero interruptions.
• Designed and implemented a DFS-based number-to-English converter, deployed it as a Python AWS Lambda function, and exposed it as a REST endpoint through API Gateway. Covered 32 unit cases, held about 200 ms median API response, and supported negatives, zero, and up to billion-scale integers.

G4F-Powered Mini Apps Suite

2025.05–2025.07

• Built three standalone Python mini apps around gpt4free, packaging provider and model selection into usable workflows across CLI utilities and a lightweight web interface.
• Developed a Game Launcher that takes a user query, requests Python game code from gpt4free, and executes the result via command line with both synchronous and asynchronous run modes.
• Implemented a YouTube Summarizer pipeline that searches videos by keyword and region, fetches transcripts by URL or ID, and generates a concise content summary from aggregated transcript text.
• Shipped a Provider and Model Visualizer as a Flask and Jinja web app on Google App Engine, exposing APIs for providers, models, refresh-providers, and generate, with basic account flows backed by Google Cloud NDB.

AI-Driven Drum Sheet Music Generation Tool Grade 11

• Developed a project aimed at integrating AI with drumming and music theory to analyze music from commercial software and automatically generate drum sheet music. While full AI-driven analysis was limited by technical and time constraints, successfully implemented a solution using ultrasonic sensors to detect drum hits and generate sheet music from live drumming performances.
• Secured a design patent for a custom microphone system developed to enhance sound detection accuracy for the project.