AiRail

• 

  Game UI Map

• 

  Comms Feed

AiRail: The Autonomous Marathon Orchestrator 🚀 Inspiration We were inspired by the transition from "Chat" to the "Action Era." Most AI tools today are static wrappers; we wanted to build a Marathon Agent—a system that breathes, plans, and executes high-fidelity infrastructure tasks over hours and days without human supervision. AiRail is a testbed for autonomous multi-LLM collaboration in a complex, localized spatial environment.

✨ What it does AiRail is a dual-agent autonomous railway simulation where AI agents don't just "chat"—they orchestrate.

Strategic Evolution: The system builds an infinite railway network, managing budgets, terrain, and city connectivity. Thought Signatures: Every action is preceded by a transparent "Inner Monologue," showing the agent's strategic reasoning (e.g., "[THOUGHT] Analyzing hub density near Tokyo..."). Autonomous Action Loop: It operates on a strict UTC-based 2-hour development cycle, simulating real-world infrastructure growth. Global Jumps: Once a city area hits a 50-station "Complexity Ceiling," the agents execute a long-distance "Jump" to establish new global hubs (London, Tokyo, Mumbai, etc.) thousands of pixels away. 🛠️ How we built it AiRail is built on a high-fidelity Orchestration Loop aligning with the Marathon Agent and Vibe Engineering tracks:

The Strategist (Gemini 3 Flash): Acts as the Senior Infrastructure Manager. It audits the world state natively using its 1M token context and generates long-term directives. The Field Engineer (Groq / Llama 3.3): Executes pixel-perfect station and track construction based on the Strategist’s plan. The Verification Loop (Surveyor): Before any construction is persistent, a Gemini-powered Surveyor verifies the proposal for logical errors or spatial collisions, forcing Self-Correction if needed. Persistence: A professional SQLite engine ensures continuity across days of execution. UI/UX: A custom Neobrutalist interface designed to WOW judges, featuring real-time signal telemetry and sub-pixel smooth infinite canvas navigation. 🧠 Challenges we ran into Aligning two different LLMs to share a single "spatial truth" was difficult. We solved this by implementing a structured Thought Signature protocol where the Engineer must justify its coordinates against the Manager's strategic intent before the Surveyor grants "Execution Rights."

🎓 What we learned We learned that the "Action Era" requires more than just prompts—it requires a robust verification architecture. By building an agent that can self-correct its own engineering proposals, we moved from a "hallucinating chatbot" to a "reliable infrastructure orchestrator."

🚀 Accomplishments that we're proud of Achieving true 60 FPS real-time telemetry on an infinite SVG canvas. Implementing an autonomous loop that has successfully run for 24+ hours, building across 5 distinct "Global Hubs" without crashing. Creating a UI that feels like a "Living Document" of AI intelligence.

🔜 What's next for AiRail Gemini Live Integration: Real-time audio/video feedback where you can talk to the Manager while she plans. Dynamic Terrain: Implementing multimodal vision analysis so agents can "see" obstacles on the map.

Built With

• canvas/svg
• express.js
• gemini-3-flash
• javascript
• llama-3.3-(groq)
• node.js
• sqlite