Max Shi
-
01
Decypherly
- Designing a novel architecture that grounds LLM reasoning in a learned domain representation, replacing the flattened text descriptions existing tools rely on
- Trained a chess encoder network to learn positional concepts while deliberately suppressing move-selection skill, forcing the LLM to use it as a reasoning primitive rather than an answer key
- Built a distillation pipeline that transfers encoder representations into the LLM; verified reliable readout of encoder outputs
- Implementing a reinforcement learning stage designed to avoid the reasoning-reward failures common to existing approaches
-
02
Faline
- Building an affordable automated litter box system that flags early signs of feline urinary and digestive illness
- Hand-fabricated hardware prototypes and designed the sensing and analysis pipeline end to end
- Partnered with 2 local animal shelters for deployment and validation
- Architecting the analysis model to explain why a health flag was raised, not just that one was
-
03
Custom Voxel Game Engine
- 3D rendering/physics engine with Metal (initially OpenGL) backends
- Render pipeline experimented with mesh shaders, LOD culling, parallelized terrain generation, and shadow mapping
- Experimented with variations to the Bounding Volume Hierarchy and heuristics (including parallel implementations) for scaling-resistant collision detection
-
04
Autonomous Trash Collection Drone Project
- Organized and led project to design and build autonomous drone for environmental cleanup
- Trained YOLO26n object detection model from scratch for edge, real-time, autonomous trash detection
- Diagnosed training plateau using data analysis; implemented Hard Example Mining to improve model performance
- Became a highlighted project at Computer History Museum Earth Day showcase
-
05
Machine Learning Intern at Manufacturing Company
- Trained CNN model to detect installation defects on high-speed assembly line with high accuracy across varied lighting conditions
- Studied image processing fundamentals (template matching, Fourier transforms) and CNN architectures to understand technical limitations
- Delivered highly accurate prototype model despite limited prior AI experience
-
06
Finetuning Image Generation Model
- Used DreamBooth to fine-tune Stable Diffusion to learn what my cat in particular looks like