Pier
R&DArchitect, Developer, AI Solutions
Multi-agent virtual studio designed as a unified work environment: a Godot 4.6 .NET desktop shell drives a registry of 31 personas (department heads, executors, utilities) backed by local Ollama models and orchestrated through Claude Agent SDK. Human brief → JSON plan → parallel execution with retry → Claude-arbitrated synthesis → report. Includes an evaluation tournament pipeline (versioned charter v1.7, multi-iteration, local LLM judge, hall of fame, JSON reports) used to benchmark, calibrate and evolve each persona against the charter. Stack: Python + Claude Agent SDK + Godot 4.6 .NET + Ollama + PowerShell. Private repo.
What this project proves
7Sail R&D · Active orchestrator + Ollama tournament pipeline
Current state
R&D • 2026
Design focus
Context
Internal 7Sail production bet: replace heterogeneous CLI scripts with a unified virtual studio, with a readable Godot UI, a native orchestrator and a continuous persona evaluation system. Target: chain shipyards (audits, productions, jam pipelines) across multiple personas in parallel without leaving the shell.
Challenges & constraints
Orchestrating 31 personas in parallel with consistency guarantees (relevant persona ↔ Ollama model mapping), retries, token monitoring, all operable from a Godot UI rather than a terminal. Building in parallel a rigorous evaluation tournament pipeline (written charter, local LLM judges, MAE / champion match metrics, hall of fame) that makes persona promotions measurable. Holding a sustainable inference cost on constrained hardware (8 GB VRAM before upgrade).
Results & learnings
Tournament charter v1.7 active since 2026-04-19, 8 iterations shipped, 0 hangs on runner v8+, local LLM judge validated (MAE 0.59, champion match 90 %), versioned hall of fame. Orchestrator core and Godot UI shell in place, Brief panel under final integration. Python + Claude Agent SDK + Godot + Ollama stack operational for studio R&D. 14B+ tuning reboot scheduled post hardware upgrade.
Tech Stack
Features
- ▸Godot 4.6 .NET shell (studio room + persona rooms)
- ▸Orchestrator: brief → JSON plan → parallel exec → synthesis
- ▸31-persona registry with Ollama mapping
- ▸Multi-iteration tournament pipeline + charter v1.7
- ▸Local LLM judge + benchmark health
- ▸Versioned Hall of Fame + JSON reports
- ▸Trunk-based architecture (feature branches < 3d)
Production note
The most structural product call was writing the tournament charter and the local LLM judge first, before the orchestrator. The charter is what makes persona promotions measurable: without it, models pile up blindly and nothing actually progresses. The Godot UI then capitalizes on those foundations, not the other way around.