Course projects / 40% seminar + 40% paper
Start from curiosity,
build a real project.
Six directions start from a runnable minimum and are designed for learning by making. Choose by interest and available resources.
Six directions
Choose by interest and resources.
Model Router Self-Learning + RSL
Build an AI gateway that selects models by task, budget, latency, and risk, then improves through independently evaluated, reversible updates.
- Good for
- Good for systems and model beginners
- Resources
- Model APIs or controlled mock services
AI Tuning of Optical Modules
Use machine learning and experimental data to find optical-module settings that satisfy bit-error, power, temperature, and stability constraints.
- Good for
- Good for data and optimization
- Resources
- Instructor data or a constrained simulator
Physical Intelligence: LiDAR 3D World Model Data Collection
Build a traceable pipeline from LiDAR point-cloud collection and synchronization through labeling, quality control, and a 3D world-model evaluation.
- Good for
- Good for data and spatial perception
- Resources
- Public data, a simulator, or approved equipment
Hardened-Weights Model Chip Design
Map a trained, fixed-weight neural network to dedicated inference hardware and measure the trade-offs in area, latency, throughput, and energy.
- Good for
- Good for chips and computer architecture
- Resources
- RTL/HLS simulators or FPGA tools
Video Generation on Low-Cost GPUs
Optimize a video-generation or video-to-video workflow for predictable memory, speed, and quality on a single consumer GPU.
- Good for
- Good for diffusion and inference optimization
- Resources
- One consumer GPU
Small Coding Model with Data Post-Training
Post-train an open coding model with traceable, permitted data and demonstrate improvement on a held-out programming task without evaluation leakage.
- Good for
- Good for model training and data governance
- Resources
- QLoRA-scale training on one consumer GPU
Project guide
Start with one small, clear question.
- ScenarioBegin with a specific user or engineering problem.
- BaselineRun a simple method before trying to improve it.
- EvidenceRecord key parameters and show change with one or two metrics.
- ReflectionShare what worked, what failed, and what comes next.
- BoundariesProtect privacy, licenses, keys, and real-device safety.
Suggested rhythm
Turn an idea into evidence in four weeks.
- 01Choose and research
Define the scenario, user, input, output, and success criterion.
- 02Minimum baseline
Build the dataset, evaluation script, and runnable version.
- 03Core experiments
Add the technical method and complete at least three comparisons.
- 04Stress test and demo
Fix failures and prepare a five-minute demo and two- to four-page report.