I'm Mark Hall (raydeStar on GitHub), a senior full-stack engineer who builds agent harnesses and evals that make AI agents measurably reliable. My focus is the part after the demo: proving an agent finishes the job, and showing why. In Sir Thaddeus, the same 1.2B local model went from 4/20 to 15/20 on a fresh verified-outcome scorecard once the harness gave it the right tools, state and checks, and any change that doesn't repeat on unseen tasks gets deleted. In HireZero, agents act only through approvals bound to the exact version reviewed, with budgets reserved before every model call. I work in Python, TypeScript and C#/.NET on Azure, and I've had fixes merged into OpenClaw and Unsloth Zoo.
Website & engineering notes · Explore the Stillwater 3D showcase · LinkedIn
A private AI assistant and agent workspace for Windows. Connect local models through LM Studio, Ollama, or another OpenAI-compatible endpoint, then work with permissioned MCP tools, durable memory, a Markdown knowledge base, and visible execution traces.
Sir Thaddeus is also my testbed for improving what a fixed model can accomplish through better tools, retrieval, and verification. The public research method documents paired experiments, repeat checks, and validation on unseen tasks.
Built with: C#/.NET, React, MCP, local LLMs.
A pipeline connecting reference images to 3D assets through Hunyuan3D, Blender, and Unreal Engine 5. It brings geometry generation, PBR texturing, retopology, rigging, deformation checks, and engine import into a workflow with review gates and traceable artifacts.
The engineering focus is reproducibility: record which reference, mesh, texture, and approval produced each asset. Human visual review remains part of the process. Explore Stillwater, the browser showcase, to see the assets in a Three.js scene.
Built with: Python, Blender, Hunyuan3D, Unreal Engine 5, Three.js.
A local-first creative workspace for planning shots, maintaining character and world references, reviewing generated images, staging 3D scenes, and assembling animated takes. Immutable media versions and explicit approvals keep creative decisions attached to the work they describe.
The architecture combines a .NET backend, a React workspace, SQLite, and recoverable provider jobs. The first Windows release is available; individual generation integrations have documented preview and validation status.
Built with: C#/.NET, React, TypeScript, SQLite, ComfyUI adapters.
An evaluation harness designed to reduce benchmark contamination and measure verified task completion. It uses fabricated evidence, sealed task banks, hashes published before measurement, and statistics that account for related test cases. It works with OpenAI-compatible model endpoints.
The public repository includes the methodology and a retired benchmark bank for inspecting and reproducing the evaluation process.
Built with: Python, paired experiments, deterministic verification.
- Grand Adventure Engine — A self-hosted AI game master and multiplayer RPG for Discord and the web. A deterministic C# engine owns combat, dice, quests, and persistent state; the language model narrates the results. Includes a WebMCP co-DM workflow with human approval for game-state changes.
- HireZero — A company of AI employees (Developer, Marketing, Support and CEO) where every push, send, publish or spend waits for the owner's approval of the exact version reviewed, and each model call reserves its budget first. Built with C#/.NET and React. Try the live demo, hosted on Vultr.
- OpenClaw: fix Workshop skill diagnostics — Fixed full Doctor lint incorrectly flagging relocated skills by resolving their paths against the original filesystem while inspecting a database snapshot.
- Unsloth Zoo: DiffusionGemma reasoning channels — Separated reasoning from answer text in the OpenAI-compatible shim, including streaming markers split across chunk boundaries.
- Keep rules, permissions, and state changes in code that can be tested.
- Make sources, tool actions, approvals, and failure states inspectable.
- Compare changes against fixed baselines and validate on unseen tasks.
- Publish limitations alongside results, and distinguish prototypes from releases.
I write about local AI, agent architecture, MCP, LLM evaluation, and trustworthy software at markbhall.dev. For professional contact, find me on LinkedIn.
Local models. Useful tools. Receipts for the magic.