Khalid Al Dosari

Fine-tuned YOLOv8s-OBB on Sixray X-ray continuously tracks and scans for predefined prohibited items in luggage belts, trained on 4 classes (Handguns, knives, wrenches, pliers), once an item is detected a voice agent contacts the relevant authority by phone call and answers any questions the dispatcher asks. Achieved mAP50 92%, precision 95%, recall 89%.

• Python
• FastAPI
• YOLOv8-OBB
• Modal
• LangGraph
• Voice Agents

Automated machine learning pipeline that takes a dataset file and the user target then runs an agent in a sandbox performing exploratory data analysis, feature engineering, data modeling and trains 13 GPU accelerated machine learning models, optimizes and benchmarks them before handing the user the wining model and its weights in addition to the cleaned dataset.

• Python
• FastAPI
• Pandas
• Modal
• LangChain
• Firecracker

Currently working on my graduation project which uses EEG brain signals to reconstruct images by using the DreamDiffusion model as a baseline and retraining a ViT-Large encoder with 75% masked signal modeling on 120K signals then aligning to CLIP on 12K ImageNet EEG pairs, reducing the number of sensors required by 84% going from 128 to 20 sensors and finally testing the results by recording new signals using a 20 sensor EEG headset to reconstruct the image the subject saw.

• Python
• EEG Signals
• Diffusion Architecture

Currently building a platform utilizing 3 agentic pipelines helping Imam University students choose elective courses best suited for them by evaluating cohort-specific prerequisites, term-specific course availability, student transcripts and GPA, finally recommending the top 5 electives best suited for each individual student.

• Python
• Agentic Workflows
• Recommendations

A dense network only vs convolutional neural networks comparison on the CIFAR-10 dataset by building a dense network from scratch and performing optimization techniques achieving a max of 59% accuracy and 1.19 plain cross-entropy loss in contrast to 91% acc. and 0.32 plain CE loss on a fine-tuned MobileNetV2, finally deploying both models for inference.

• Python
• PyTorch
• TensorFlow
• Modal

Developed an AI-powered tourism analytics platform for the PwC Empowerthon, securing 4th place out of 11 teams. Integrated a bilingual analytics dashboard and ROI calculator using React, TypeScript, Vite, Supabase and JavaScript, aiding Saudi destination owners turn visitor data into actionable insights.

• React
• TypeScript
• Vite
• Supabase
• Voiceflow