Senior Software Engineer with 7+ years building full-stack enterprise systems at scale, now going deep on AI engineering.
I architect C#/.NET applications that process trillions in historical data. I'm channeling that enterprise engineering discipline into modern AI: RAG pipelines, multi-agent workflows, MCP tooling, and local LLM inference on my NVIDIA DGX Spark.
SparkBench — Benchmarks and fine-tuning studies on my NVIDIA DGX Spark (GB10, 128 GB unified memory). The headline experiment asks one question: can a LoRA fine-tuned 7–14B open model, running locally, replace the frontier LLM inside AgentReview's Quality Agent at comparable precision/recall and near-zero marginal cost per review? Three conditions — frontier baseline, untuned small model, LoRA-tuned small model — run through the same frozen harness, trained with Hugging Face peft/trl, served with vLLM and Ollama, and tracked end to end with MLflow and DVC. The methodology is pre-registered: all three possible outcomes are declared publishable up front, training is capped at five full runs, the held-out test set is evaluated exactly once, and results ship with bootstrap confidence intervals. Also includes a reproducible inference benchmarking suite measuring throughput, latency, and memory across models and quantization levels.
DriftWatch — An end-to-end MLOps pipeline that predicts equipment failure from sensor time-series and, more importantly, stays healthy after deployment, with a dashboard now live at driftwatch.vondraysanford.com. MLflow experiment tracking and registry, DVC-versioned data, GitHub Actions CI/CD to an Azure ML endpoint, and Evidently drift monitoring wired to an automated retrain trigger. The thesis: the model is 20% of an ML system — this project is the other 80%.
AgentReview — Multi-agent code review in C#/.NET 10, demo now live at agentreview.vondraysanford.com. An orchestrator fans a PR diff out to quality, security, and docs agents that call real tools (Roslyn, Semgrep, GitHub) through MCP, then synthesizes one ranked review. Measured results: 100% precision and human agreement, 17/18 planted-bug recall, $0.068 per review, all traced with OpenTelemetry. Proof that sophisticated agentic patterns work reliably in the .NET ecosystem — not just Python.
DocQuery — A local-first RAG application for querying documents in natural language, built in public with C#/.NET 10 and React. Phases 1–4 complete with a demo now live at docquery.vondraysanford.com; Phase 5 turns it into a daily study tool.
The 10X Engineer Toolkit — A stack-agnostic library of engineering practices packaged for AI coding agents, installable as a Claude Code plugin and portable to any assistant. Skills, agents, prompts, templates, workflows, and configs that codify how great engineers plan, review, and ship — so every ticket, review, and deploy gets the same care automatically.
Twitter Sentiment Analysis Bot — A shipped Python ML pipeline that classifies tweet sentiment and detects automated accounts with an XGBoost model (0.89 ROC AUC on 10K+ accounts), served through an interactive Discord bot. Includes RSA-signed model integrity checks and a full collection-to-deployment data pipeline.
- MonoGame (14K+ ⭐) — Authored XML API documentation for the
GraphicsAdapterandAlbumclasses across 2 PRs — both merged (Aug 2026) - Cataclysm-DDA (13K+ ⭐) — Fixed the solar cell's implausible half-meter
longest_sideand addedlooks_likesprite fallbacks for the Xedra Evolved dream weapons across 2 PRs — both merged (Aug 2026) - Kana-Dojo (3.2K+ ⭐) — Standardized the Japan trivia answer format to support multiple accepted answers and added a「〜ても」grammar entry across 2 PRs — both merged (Aug 2026)
- KodeKloud AI-102 — Submitted a security fix replacing hardcoded Azure credentials with placeholders in a public course code sample (open, awaiting review)
Languages: C#, TypeScript, JavaScript, Python, T-SQL, HTML/CSS
Frameworks: .NET Core, ASP.NET, React, React Native, Node.js, Entity Framework
Databases: SQL Server, PostgreSQL, MySQL, Redis
Cloud & DevOps: Azure, Docker, Kubernetes, CI/CD, Azure DevOps, Argo
AI & ML: Azure AI Services, Ollama, RAG Pipelines, LLM Integration, Claude Code, GitHub Copilot, MCP, Prompt Engineering
Hardware: NVIDIA DGX Spark (local inference and fine-tuning)
| Certification | Issuer | Date |
|---|---|---|
| Azure AI Engineer Associate (AI-102) | Microsoft | Jun 2026 |
| GitHub Copilot (GH-300) | Microsoft | Jun 2026 |
| Azure Fundamentals (AZ-900) | Microsoft | Jun 2026 |
| GitHub Foundations (GH-900) | GitHub | Jun 2026 |
| Azure AI Fundamentals (AI-900) | Microsoft | Apr 2026 |
| Claude Code in Action | Anthropic | Jun 2026 |
| Introduction to Model Context Protocol | Anthropic | Jun 2026 |
| Introduction to Subagents | Anthropic | Jun 2026 |
| Introduction to Agent Skills | Anthropic | Jun 2026 |
📖 Machine Learning Operations Engineer Associate (AI-300) — In Progress
- Studying for the AI-300 (Microsoft Machine Learning Operations Engineer Associate) — operationalizing ML and generative AI solutions on Azure
- Running local LLM inference and fine-tuning experiments on DGX Spark via SparkBench
- Sharing what I learn at vondraysanford.com and on LinkedIn