Local models. Real tools. Honest measurements.
By day, I build production software with .NET and Azure. This profile is my workshop for local AI, agent tooling, game systems, and experiments in making smaller language models genuinely useful.
Most of my career has been spent building C# systems. Lately, I have been exploring the unglamorous machinery that turns an impressive AI demo into dependable software: deterministic execution, Model Context Protocol (MCP) integrations, explicit permissions, retrieval-augmented generation (RAG), durable memory, visible evidence, human approval around destructive actions, and repeatable evaluation.
Project notes, experiments, and longer-form writing at markbhall.dev
An open-source, local-first AI assistant and agent workspace for Windows. Sir Thaddeus connects local models through LM Studio, Ollama, or another OpenAI-compatible endpoint, then gives them permissioned MCP tools, durable memory, private knowledge, visible sources, execution traces, routines, and an emergency stop.
It also serves as an evidence-driven research harness for measuring how deterministic tools, retrieval, memory, permissions, and verification improve the real-world task performance of fixed small language models.
C# · .NET · Avalonia · MCP · Local LLMs · RAG · Agent evaluation
A self-hosted multiplayer AI-narrated RPG for Discord and the web. The language model tells the story, while a deterministic C# game engine resolves combat, dice rolls, loot, quests, inventory, and skill checks. The result is generative storytelling with rules that remain fair, testable, and consistent.
C# · .NET 10 · ASP.NET Core · SignalR · Docker · Playwright · LM Studio · Ollama
- Unsloth Zoo PR #769: Added reasoning-channel extraction for DiffusionGemma's OpenAI-compatible shim, including streaming boundary handling and automated tests. Merged upstream into the Unsloth ecosystem.
- Keep the LLM out of the critical path when correctness matters.
- Use deterministic systems for rules, state transitions, validation, and destructive actions.
- Give agents bounded capabilities, explicit permissions, visible evidence, and a hard stop.
- Measure improvements against frozen baselines and unseen tasks, not impressive-looking demos.
At markbhall.dev, I publish engineering notes and experiments on local language models, agentic AI, MCP tooling, LLM evaluation, benchmark design, production AI safety, and AI-assisted software development.
Agentic AI · Model Context Protocol (MCP) · Local LLMs · LLM evaluation ·
RAG · Tool use · Structured outputs · Human-in-the-loop AI · .NET / C# ·
Azure · Docker · Playwright