Cloud Computing
Levon Blog · Levon Zhao
I Focus on Game & AI.
Part-time AI Fellow at Handshake AI, evaluating frontier models. Previously built anomaly detection and agentic analytics at Activision Blizzard.
My personal blog and portfolio as an AI / ML engineer, with notes on AI, computer graphics, and the systems behind games and software.
Professional credentials
Cloud · Data & AI · Deployment · Finance
Data & AI
Deployment & Finance
Selected work
All projectsMicrosoft Gaming AI Project
At Activision Blizzard, built a PB-scale anomaly-detection and agent-assisted investigation system for 50+ live-service KPIs, reducing analyst response latency by 40%+. Productionized daily ML workflows from detection to reporting.
- Databricks
- Airflow
- RAG
- Agentic AI
- ML Workflow
YOLO-KAN Research
First-authored an IEEE CAI 2025 short paper integrating KAN modules into YOLOv11n for Microsoft COCO object detection. The best KAN-2-5 configuration reached 65.83% precision (+1.84 pp over the 63.99% baseline) while reducing network depth from 319 to 299 layers.
- YOLOv11
- KAN
- PyTorch
- Microsoft COCO
- IEEE CAI 2025
AI Companion
LIVEAn English, source-linked voice and text companion that follows visitors across the portfolio and answers focused questions about my experience, research, and projects.
- Next.js
- LiveKit
- Gemini
- Voice AI
- Source-linked
Education
Georgia Tech MSCS (AI)GPA 3.88 / 4.00
NVIDIA Sponsored Agentic AI · Artificial Intelligence · Game AI · Game Design & Development · GPU Hardware & Software · Machine Learning for Trading
Degree audit (PDF)Northeastern MSCSGPA 3.97 / 4.00
Data Visualization & GenAI · Distributed Systems · Cloud Computing · Deep Learning Research
Unofficial transcript (PDF)Latest from the blog
View all articlesDeferring KIRA’s 3D Avatar Until the First Conversation
How two React states defer KIRA’s WebGL avatar until first use, preserve it after closing, and make its lifecycle observable in the browser.
Debugging a Beier-Neely Image Morph
What 26 line pairs, five transformation studies, and several coordinate bugs taught me about image warping.
YOLO-KAN: What the Ablation Experiments Taught Me
Introducing Kolmogorov-Arnold Network modules into YOLO11n and testing the architectural trade-offs.