AI Systems

Agentic AI, built like production software

AI systems I've designed and built end to end: 10+ production systems for a game team, a multi-agent platform, a business AI product, and an AI co-pilot inside a professional 3D application. All of them treat the model as one part of a system that has to be tested, permissioned and observable.

Patterns I use

Agents & tool calling

Assistants that delegate to specialist agents and call real tools: Maya, Unreal, Miro, Monday.com, file systems, code execution.

RAG & memory

Vector embeddings and retrieval over project data; per-user and per-org memory that consolidates on a schedule.

Cost-aware routing

Tiered model routing: cheap models gather, stronger models answer. Claude, OpenAI, Gemini and local Qwen behind one interface.

Guardrails & permissions

The engine owns facts and permissions, not the model. Tools carry safety levels; private data needs consent before it's shared.

Evaluation

Agent output is judged by independent code checks, not by another model's opinion. “Done” has to be proven.

Sandboxed execution

Agent-written code runs in a no-network Docker sandbox with restricted files.

Self-hosted inference

Local LLMs and Stable Diffusion on my own GPU workstation, exposed through Cloudflare tunnels.

Integrations

MCP servers, WebSockets, Discord bots, a Qt desktop client and web dashboards on Flask and FastAPI.

Production AI · Rascal GamesAI shipped inside a game team

At Rascal Games I built 10+ AI systems with Claude for an unannounced Unreal Engine 5 title.

  • AI-driven characters: NPCs that decide for themselves: interacting with objects in the world such as arcade machines, choosing adventures to go on, and picking their own clothing.
  • Item catalog management: an AI system that managed gear items and added new ones into the game's item ecosystem.
  • Automated setup: a system that builds Unreal animation blueprints automatically and sets up AI agent characters in the game.

The same patterns apply outside games: agents making decisions inside rules, AI maintaining a product catalog, and code generation that removes repetitive setup work.

Business AI productNotera

A multi-tenant AI platform for production teams: recorded meetings become decisions, documents and tasks. A central assistant orchestrates domain agents for minutes, docs and tasks.

  • Multi-tenant onboarding, meeting transcription, Monday.com integration.
  • Column-level encryption for customer data; real-time streaming over WebSockets.
  • Next.js, FastAPI, PostgreSQL, vector embeddings, multi-model orchestration.

AI inside a professional toolMaya co-pilot

An assistant built into Autodesk Maya that knows the application and my codebase, and runs tools on demand. It routes between Claude and OpenAI by task and classifies generated code by safety level before running it.

Maya AI assistant answering a question about the codebase

Connecting agents to softwareMCP servers

  • Maya: scene editing plus a full facial-rig workflow an agent can drive.
  • Unreal Engine: actors, Blueprints, materials and Control Rig, with screenshots to verify results.
  • C++ plug-ins: 32 tools generated from one spec, so the agent layer can't drift from the API.

Multi-agent platformBot Town

A platform where AI agents with distinct roles (producer, reviewer, art director, coder) propose, build and review work together. A personal assistant and an org assistant share information through a relay that asks the user before anything private leaves.

Bot Town architecture Chat surfaces feed two assistants, Notera and Hera, who delegate to worker agents. The engine owns tasks, artifacts and permissions in SQLite. Code runs in a no-network Docker sandbox, using local and hosted models. SURFACESASSISTANTSWORKERSENGINEEXECUTION Discord Desktop app (Qt) Web control room Godot desktop pet Noterapersonal assistant Heraorg assistant · reviewer relay (asks before sharing) Mira · art direction Forge · code Pixel · character art + hired workers State & contracts tasks · artifactsreviews · memorytyped handoffsaction permissionsSQLite SandboxDockerno networkindependent checks Modelslocal Qwen (GPU)Gemini · Claudetiered routing The rule that shaped it: an agent saying “done” isn't evidence.The engine records acceptance, the artifact, the checks and the review as separate steps.
  • “Done” isn't evidence: task acceptance, artifact, checks and review are recorded as separate, auditable steps.
  • Typed contracts: after a model reviewer approved a misleading proposal, handoffs moved to code-checked input/output contracts.
  • Stack: Python, Flask, SQLite, Docker, PySide6, Discord; local Qwen and Stable Diffusion; Gemini, Claude and OpenAI.

The coding trials are supervised, and described that way on purpose.