Projects

UnityMCP is an open-source AI agent platform that lets AI assistants like Claude, Codex, and Cursor operate inside the Unity Editor. Most competing MCP servers at the time had no safety layer — UnityMCP was designed safety-first, prioritizing guardrails, evaluation, and agent reasoning alongside its 45+ built-in tools.

Safety Guardrails

Every tool carries annotations — destructiveHint, readOnlyHint, idempotentHint, openWorldHint — that inform AI assistants about the risk level of each operation before execution. A safety blacklist prevents AI from executing dangerous editor menu items. These guardrails were a deliberate design choice to make users comfortable trusting AI with their Unity projects.

Agent Reasoning Architecture

Four workflow prompts guide AI assistants through multi-step reasoning tasks: inspecting GameObjects, examining prefabs via prefab stage, safely modifying components with verification, and setting up new scenes with appropriate defaults. A progressive discovery pattern via search_tools lets the AI explore available capabilities incrementally by category or keyword rather than receiving all tools at once — shaping how agents reason about tool selection and sequencing.

Observability & Monitoring

A real-time activity log monitors all MCP requests and responses from the editor window. Built-in test tools (test_echo, test_add, test_unity_info) validate connectivity and parameter handling.

Model Evaluation

Tool-calling behavior was tested and iterated across Claude 4.5 and Claude 4.6. Tool design, parameter schemas, and prompt templates evolved based on how different model versions interacted with the platform.

Developer Extension API

A single-attribute registration system — [MCPTool], [MCPResource], [MCPPrompt] — lets developers extend the platform with custom tools, resources, and prompts without modifying core code. Parameter constraints (enums, min/max) are included in JSON Schema sent to AI assistants. Automatic discovery on domain reload means no manual registration needed, making UnityMCP a platform for the community.

Architecture

A native C plugin runs the HTTP server on a background thread, persisting across Unity domain reloads — the AI connection stays active even during script recompilation. JSON-RPC over HTTP(S) with optional TLS encryption and API key authentication supports remote access. Cross-platform: Windows, macOS, Linux. Open source under GPLv3.

Developed brand strategy recommendations for the Cheetos portfolio at Frito-Lay. Applied consumer insights and market analysis to identify growth opportunities and inform positioning decisions within the broader PepsiCo snack brand ecosystem.

Evaluated Truist’s digital presence across credibility, consistency, storytelling, reach, and visibility. Delivered a 10-minute presentation with actionable recommendations for driving traffic and improving brand perception online.

Working directly with the head product manager on core platform decisions at Form.io. Owning workstreams within a multi-phase portal UI/UX modernization initiative aimed at reducing onboarding friction, streamlining developer workflows, and enabling new platform capabilities.

Approach

The initiative uses a “tik-tok” prioritization approach, alternating major and minor efforts to maximize development throughput while maintaining momentum across multiple workstreams.

A/B Testing & Data-Driven Decisions

Running A/B tests on UI changes for platform features, defining test parameters, and analyzing results to inform product direction. Each phase of the initiative is validated with data before moving forward.

Strategic consulting engagement developing US market entry, subscription pricing, marketing playbook, and 6-12 month implementation roadmap for PIBOCO (Picture Book Collective), a Copenhagen-based interactive digital publishing platform.

Conceived NuSol as an original venture idea and took it from concept through prototype development. Conducted market validation, consumer research, and delivered a consulting-style pitch with actionable positioning strategy.

Conceived and pitched Vouch, a peer-to-peer character reference platform. Developed the concept from ideation through prototype, built a financial model, validated the market opportunity, and delivered a final investor-style pitch.

Applied the Dreams and Details framework to analyze Adobe’s response to generative AI. Examined their commercial safety strategy (training only on licensed content), platform ecosystem integration into Creative Cloud, and enterprise indemnification approach that turns ethical AI into a competitive moat.

Reflex was the scene-scripting backbone at Blink Studios. It let our designers lay out camera cuts, character movements, and dialogue triggers on a visual timeline inside Unreal, then play them back in-engine without writing any C++. The implementation is proprietary, but it cut our scripted-scene production time dramatically.

Specular is a tree-based dialogue scripting language I built at Blink Studios. It compiles down to a bytecode format that Unreal Engine reads at runtime, so our writers could create branching conversations, trigger game events, and do math without touching C++. I wrote the grammar in ANTLR4 and the transpiler in Java. The Unreal plugin that executes Specular scripts is closed source, but the language itself is on GitHub.

Contributed the Python3 implementation of antlr-denter, a library that generates Python-style INDENT and DEDENT tokens for ANTLR4 grammars. Without it, writing indentation-sensitive languages in ANTLR means building your own token injection layer from scratch. I needed this for my own work building Specular’s grammar tooling and ported the existing Java implementation to Python3 so the broader ANTLR community could use it too.