Neuro Symbolic Game Level Designer

Problem. Building a 2D isometric game level is slow, manual work. Designers slice spritesheets by hand, tag every tile (floor or wall? walkable? where is its anchor?), paint maps tile by tile in an editor, and then write engine code to load them. Asking a generative AI to "make a level" does not work either: models that output raw tile grids drift, invent tiles that don't exist, and produce maps you cannot walk through. Solution. Neuro Symbolic Game Level Designer is a web app built on one principle: AI agents plan, algorithms build, a validator guarantees. A developer types a prompt such as "haunted graveyard with a woodcutter's camp and a boss arena in the north", optionally uploads their own isometric spritesheets, and gets a playable level in about a minute. It runs in three pipelines: 1. Data Ingestion: algorithms detect the tile size and slice the sheet. Then four specialist AI agents (boundary, classification, collision, entity) inspect every tile in parallel, using strict, validated schemas. A harmonizer merges their answers in code and calls AI only where the agents disagree. The result is a semantic asset catalog, cached for instant reuse. 2. Level Planning: a Topology Agent plans a room graph (rooms, directions, corridors, enemies, style weights), never raw tile arrays. Deterministic algorithms then turn it into an exact isometric grid: they place the player, exit and zombies, dress floors and props, and seal the level with a natural hedge. A validator checks every rule (path from spawn to exit, every room reachable, no escape routes). Failures go back as structured errors in a self-healing loop, as a new layout or a new plan. 3. Execution: an AI agent designs enemy behavior per room (patrol, guard the exit, idle until the player is near) as data, never as code. Compilers write a standard Tiled isometric map with web-safe texture atlases, and templates generate typed Flame (Dart) loader code. A verification engine runs 6 checks, including a real dart analyze on the generated code. Target users and interaction. The users are indie developers and game jam teams using the Flutter Flame engine. They type a prompt, watch each pipeline step in a live progress view, preview the map, and press Try Out to play it immediately in the browser. The game has zombies with their designed behaviors and a health system. They can also download a bundle (map, tilesets, Dart loader, verification report) to drop into their own game. What makes it unique. The AI is used only where judgment matters: understanding art, planning a level, designing enemy behavior. Everything that must be exact is done by deterministic code, so levels are correct by construction and verified, not guessed. It works on any isometric spritesheet, not one hard-coded tileset: we tested it on a grassland pack, a desert pack and Kenney farm and library sets. Its outputs are open standards (Tiled maps) plus code a developer owns. Efficiency. Mechanical work uses no AI tokens. Ingestion results are cached, so a known sheet costs zero AI calls. Only one AI call plans the level, and one designs its enemies.