Dual LLM, Vendor Agnostic Architecture
The clue generation layer is completely provider independent. Switch between AI vendors with a single environment variable, zero code changes required.
Default production provider. Cost optimised with a clue caching layer, identical words are never clued twice, cutting API spend dramatically on large runs.
Fully integrated as a hot swap alternative. Uses the OpenAI compatible Gemini endpoint, the same prompt and validation stack applies to both providers.
The pipeline stops and reports gracefully when API credit is exhausted, protecting against runaway costs and leaving no partial puzzles in the output.
7 Point Quality Gate
Every grid is interrogated against seven hard constraints before it proceeds to clueing. Failures are discarded, not patched.
| # | Check | Threshold | Why It Matters |
|---|---|---|---|
| 01 | Minimum Word Count | ≥ 40 words | Ensures standard puzzle density. |
| 02 | Zero Duplicates | 0 allowed | No repeated answers within one puzzle. |
| 03 | Short Word Cap | ≤ 25% 3 letter words | Eliminates cheap filler grids. |
| 04 | Letter Diversity | ≥ 18 unique letters | Guarantees a healthy alphabet spread. |
| 05 | Intersection Integrity | 100% match | Every Across/Down crossing must agree. |
| 06 | Word Length Bounds | 3 to 15 characters | Classic crossword conventions enforced. |
| 07 | Clue Self Reference Block | 0 allowed | No clue may contain its own answer. |
From Blank Grid to Published Puzzle
Every puzzle flows through this automated workflow before it reaches a player's device.
rotational symmetry
300K+ validated word list
no partial passes
with clue caching
to GCS & served via CDN
Smart Metadata & Bulk Scale
The pipeline classifies, names, and deploys puzzles, and is engineered to run unattended for days on a cloud server.
-
Auto Difficulty Classifier
Easy / Medium / Hard is assigned automatically based on average word length and obscurity density. No manual tagging ever required.
-
AI Puzzle Titling
The LLM analyses answer themes and generates a creative title per puzzle, e.g., "Grocery Store Groove," "Lunar Legends," "Sleigh Trackers."
-
Resume Safe Bulk Mode
Scans existing files for the last puzzle ID and continues from there, after any interruption (timeout, credit exhaustion, server restart), with no data loss.
-
Incremental Cloud Sync
Each puzzle is written to disk and synced to GCS immediately. The master index regenerates automatically, live app installs discover new puzzles on their next load.
Browse Real Puzzles From the Library
These are actual puzzles produced by the generator, grids filled algorithmically, clues written by AI, quality validated through the pipeline.
Interested in the Generator?
The generator is included in the full acquisition. Reach out to discuss terms.
← Return to Overview