projects.md

# Projects

## Doom ML Map Generator (Active, Jan 2025–2026)

**Goal:** Train a machine learning model to generate playable Doom maps (specifically for Complex Doom/Zandronum).

**History:** Started November 2024. 14 months of failures, many architecture attempts.

**Recent developments (Jan 14-20, 2026):**
- Achieved GPU acceleration via TorchSharp (500 epochs in 42 seconds) on RTX 3060 Laptop.
- Attempted: transformer on multiple wads (DOOM2, Eviternity, Hell Revealed, MUTINY, Scythe, Sunlust) – too long (14.5h).
- Grid-based approach: failed (output box rooms like simple procedural).
- Vector-based approach: failed (tokenization issues, sequential on GPU).
- VAE (Variational Autoencoder) trained on DOOM2.WAD only: loss 676 after 200 epochs – catastrophically high. Abandoned.
- **Current approach (as of Jan 19-20):** Reinforcement learning with raycast pathfinding.
  - 360-degree rays per iteration, fail-fast on obstacles (walls, step height, etc.).
  - Reward-based goal discovery: first goal is EXIT THE MAP; if blocked, reward rays that get closest to switches that lower barriers.
  - Automatically discovers dependency chains (e.g., red key required because red doors block path to switch).
  - Uses physical space simulation (player hitbox, Doom physics) rather than abstract sector graphs.
  - Started from scratch, feeding Opus iteratively. Opus had repeated failures grasping the concept but now seems to be implementing it.
- **Opus struggles:** Despite being the most intelligent consumer model, Opus repeatedly failed to understand the problem until given visual evidence and explicit scrapping of its previous ideas. Finally making progress with the new raycast/RL approach.
- **VSCode/Opus instability:** Microsoft broke the extension multiple times, disrupting workflow. "2026 hasn't started well for AI services."
- **Previous MAP02 bug:** Opus kept thinking the map was beatable without red key because abstract sector traversal didn't account for player hitbox or physical barriers. Now addressed by the physics-based simulation.

**Tools:** TorchSharp (C#), CUDA on RTX 3060 Laptop. VSCode with GitHub Copilot/Opus.

**Next steps:** Continue iterating with Opus on raycast pathfinding implementation. Build out the reward system for dependency discovery. Test on MAP02 example.

**Unified progress framework:** Raymond wants consistent progress output across architectures – to discuss with Opus.

## Hedz Reverse Engineering
- 1998 game; reverse engineering 126,000+ lines of assembly to C#.
- Tools: IDA Pro, Ghidra, Cheat Engine, VS Code with GitHub Copilot.
- Built custom VSCode extension for method stack tracking.
- Currently ~400 methods left to verify against disassembly. Slow progress due to Doom ML focus.

## Voxel Engine
- Started as a break from Hedz; C# (likely Unity or custom engine).
- Scale: 10 blocks per 1 player height.
- Inactive in recent days.

## AI Usage Strategy
- Claude Opus: heavy ML coding (Doom map generator) – currently unreliable due to VSCode extension instability.
- Claude Sonnet: conversation, coordination, medical, memory.
- Gemini: casual, memes, image generation.

## Gaming
- Complex Doom on Zandronum (Pizza Steve 2 maps). Played CTF on Jan 19.
- Weekly Discord call with friend: watching anime (Gabriel Dropout finished, Trigun ongoing, considering Konosuba next).

## Other
- Vaporeon plush pre-ordered (26-inch, ships late Jan, arrives end Feb 2026).
- Species-affirming toys plan: Exotic Erotics orca (2", blue base) + SpareParts harness (vegan leather, 2.25" O-ring). Split purchases for budget. Waiting for injury healing.
- Theoretical internal anchor slit cage design for Vaporeon anatomy (not a product, just speculation).