Self-driving Car Problem · Tran Quoc Truong

Why this case is included

It shows the early applied-AI foundation before the LLM/backend work.

This case should not replace Omni-Agent or APIT as the main story. It is included as earlier evidence: an award-winning research paper that shows computer vision, simulation, model training, control logic, and technical writing experience.

System

Unity simulation, YOLOv8 perception, and PID control.

SimulationUnity environment with lane markings, stop lines, signs, shadows, snow, and obstacles.

Traffic signsColor filtering crops signs, then YOLOv8 Classifier predicts seven sign classes.

Lane segmentationYOLOv8 Segmentation detects lane masks from VIA and self-labeled data.

ControllerLane midpoint and PID-style control steer the vehicle through the course.

Data and result

The result is useful, but framed honestly.

The traffic sign task used 18k+ images across seven classes. The lane segmentation task used roughly 8k provided images and 1.5k self-labeled images from the simulation environment, with augmentation.

The system completed the simulation benchmark in 125.8 seconds with full marks. The paper also clearly documents limitations: color-filter robustness, simulation-only evaluation, and primitive controller assumptions.

Award proof

Second Prize evidence.

The award image is included as proof, but this case remains in the "Earlier AI Research" tier so the portfolio does not drift away from Applied AI systems, agent evaluation, and backend control-plane work.