Hands-on projects on building AI agents and on training, serving and measuring the models behind them, mostly on Azure
This repository collects the hands-on work of Xinyu Wei (魏新宇), Microsoft AI Lead Senior System Engineer, in two connected areas:
- AI agents: building, hosting and operating agents on Microsoft Foundry and Azure, the model APIs and gateways they call, voice agents, retrieval and MCP.
- Deep learning: fine-tuning and reinforcement learning, inference optimization, precision and quantization, GPU selection on NVIDIA and AMD hardware, multimodal models, and one-topic algorithm explainers.
Each folder is a self-contained project with its own README. Many recent projects also include the scripts, raw results and CI checks that produced their numbers.
| Last updated | Project | Area |
|---|---|---|
| 2026-11-14 | Thermal 3D Modeling Studio: Azure OpenAI GPT-6 Astra Writes the Model Code, OpenCascade Builds the STEP | AI Agents |
| 2026-11-05 | Component Precision and Deployment Guide for Large MoE Models | Deep Learning |
| 2026-11-03 | Decision Model on the Edge: Fine-Tuning and Inference Optimization | Deep Learning |
| 2026-10-27 | Claude 5.5 on Microsoft Foundry: Thinking, Effort, Prompt Caching, Latency and Coding Accuracy | AI Agents |
| 2026-10-23 | Five Ways to Host an Agent Sandbox on Azure | AI Agents |
| 2026-10-17 | LLM Inference Optimization on Azure GPU VMs | Deep Learning |
| 2026-10-13 | Azure MI300X: AllReduce regression after a native host-driver upgrade | Deep Learning |
| 2026-10-11 | Measure Azure GPU Hours with DCGM and Azure Monitor | Deep Learning |
| 2026-10-04 | Azure OpenAI Model and Model Router Benchmark | AI Agents |
| 2026-10-04 | English meeting transcription on far-field audio: six Azure speech paths, four AMI clips | Multimodal Models |
| Area | What you will find |
|---|---|
| 🤖 AI Agents | Building, hosting and operating agents on Microsoft Foundry; agent applications; voice agents; model, API and gateway benchmarks on Azure; retrieval and evaluation |
| 🧠 Deep Learning | Fine-tuning, RL and alignment; inference optimization (vLLM, SGLang, speculative decoding, KV cache); precision and quantization (FP8, FP4); GPU selection and benchmarks on NVIDIA and AMD; distributed training |
| 🔬 DL Algorithm Insights | One algorithm or systems topic per folder, explained in depth with experiments |
| 🎨 Multimodal Models | Image generation benchmarks (MAI-Image, GPT-Image), speech and voice, vision-language models, medical imaging |
| 📘 Edge-Side AI book code | Companion code for Edge-Side AI: Training, Optimizing and Deploying Small Models (2026) |
| 📗 LLM book code | Companion code for Principles, Training, and Applications of Large Language Models (2024) |
Each area page groups its projects by topic. Projects whose models or APIs have since been superseded stay in place under Archive on their area page, so existing links keep working.
Agents & platforms: Microsoft Foundry • Azure OpenAI • Microsoft Agent Framework • MCP • Azure AI Speech / Voice Live
Training & serving: vLLM • SGLang • DeepSpeed • FSDP • LoRA / QLoRA • GRPO
Infrastructure: Azure ND / NC GPU VMs (H100, H200, MI300X) • AKS • InfiniBand / RDMA
"Edge-Side AI: Training, Optimizing and Deploying Small Models"
- 🔗 Repository: Code Examples
- 🛒 Purchase: JD Mall
- 🔗 Repository: Code Examples
- 🛒 Purchase: JD Mall
- 🔗 Repository: FSI-IT-Construction
- 🔗 Repository: MSA-DevOps
- 🔗 Repository: OpenShift Applications
The full repository is large (over 3 GB). To download only the folder you need:
git clone --depth 1 --filter=blob:none --sparse https://github.com/david-xinyuwei/david-share.git
cd david-share
git sparse-checkout set Deep-Learning/Speculative-DecodingReplace Deep-Learning/Speculative-Decoding with any project folder. Videos, audio and some large result files are stored with Git LFS; if a project needs them, install Git LFS and run git lfs pull --include="Deep-Learning/Speculative-Decoding/**".
This project is licensed under the MIT License - see the LICENSE file for details.
⭐ Star this repository if you find it useful.