PROJECT DETAIL

Croztic

An acrostic puzzle game built with AI assistance from top to bottom, from the gameplay itself to a pipeline that generates thousands of puzzles for pocket change.

PROJECT SUMMARY

A satisfying puzzle where every letter you uncover reveals so much more.

Croztic is an acrostic puzzle game I architected and built with agentic AI in under a week. Beyond the game itself, I wrote the specs, designed a 4-stage AI pipeline that generates puzzles from trivia and word pools, and operated AI to create both a webapp and in-game admin tools for puzzle auditing, all connected via API.

Acrostic puzzle game AI-assisted development 4-stage content pipeline Cross-platform tooling

Gameplay Video

Game Development Accelerated by AI

Built a satisfying acrostic puzzle accelerated by Agentic AI

What I Built

Spec building with AI assistance to ensure efficient assisted development and appropriate design decisions.
Built the entire gameplay system from the ground up: the acrostic board, clue-to-letter mapping, the cascading reveal logic, and all supporting UI, with agentic AI accelerating everything from initial scaffolding to final polish.
Designed the interaction flow so every correct answer feels rewarding: letters populate across the board and previously impossible clues suddenly become solvable.
Focused on tight UX feedback: snappy animations using Tween tools and animators, clear visual cues when letters slot into place, and zero ambiguity about what changed after each answer.

Impact

Created a gameplay loop where progress compounds: solving one clue literally makes the next one easier, keeping players in flow.
Shipped the full game, mechanics, UI, and polish, in under a week, with agentic AI handling boilerplate, iteration, and polish passes so development stayed focused on design decisions.
Demonstrated how AI can accelerate development timelines without sacrificing quality, by handling everything from scaffolding to iteration to polish across the entire project.

AI CONTENT PIPELINE

Architected a 4-stage AI pipeline.

What I Built

Designed a multi-stage AI pipeline that generates puzzles through four coordinated passes, each with a distinct responsibility:
Stage 1, Category & Difficulty: AI produces a batch spec, e.g. "levels 1000–1200", with categories and difficulty tiers, output as category-to-difficulty.json.
Stage 2, Maintext & Trivia: A second AI pass reads that spec and generates the main phrase and trivia text for every puzzle in the batch.
Stage 3, Word Pools: Separately, AI generates 500 easy, 500 medium, and 500 hard word pools. Each pool contains clues organized by sub-category: trivia, wordplay, definition, synonym, antonym, and more. The difficulty tier determines which types and how many of each get pulled for a given puzzle.
Stage 4, Assembly: Here the meta layer: AI generates the Python script itself, the tool that merges all three inputs (difficulty spec, trivia text, and word pools) into fully assembled, validated puzzles in seconds.
Connected the pipeline to a companion webapp and in-game admin mode via API, so puzzles flow from generation to audit to deployment without manual handoffs.
Designed the system to be fully modular: each stage can be re-run independently, word pools can be regenerated for different difficulty curves, and the assembly script can be re-generated if game design rules change.

Impact

The pipeline can produce roughly 1,000 puzzles a day for $1–2 in API costs, with automated difficulty balancing built into the architecture.
Puzzle designers and developers can create, audit, and deploy content from anywhere, whether browser, Unity editor, or in-game admin, without waiting on anyone else.

Collaboration Tools

Created Collaborative Tools with LLM AI - Game build

What I Built

An in-game built in admin-mode that enables real-time clue modification that syncs with cloud database.
An API end points that syncs up the game, AI pipeline, and webapp editing tools.

Impact

Enable seamless collaboration between developers producers and testers when they try to feel the game in the build.
Improving puzzle quality through real-time feedback and iterative adjustments.

Collaboration Tools

Created Collaborative Tools with LLM AI - Editor and Webapp

What I Built

An editor utility tool for auditing puzzles and words that syncs with cloud database without playing the game.
An API end points that syncs up the game, AI pipeline, and webapp editing tools.

Impact

Enable full control over puzzle content and game mechanics through the admin tools in every platform possible.
Enable seamless collaboration between developers and puzzle designers without needing to touch developer engine and or game build.

OPTIMIZATION & ARCHITECTURE

Kept things lean, deliberately.

What I Built

Optimized the WebGL build for fast loading and smooth runtime.
Designed the system modularly: the game, the AI pipeline, and the auditing webapp are all decoupled and communicate through clean interfaces, so any piece can evolve independently.
Made deliberate architectural calls for puzzle loading while keeping scale in mind.

Impact

Achieve lightning-fast load time even on low-end devices, achieving under 5 seconds for initial load.
Maintain a small memory footprint, with 3,000+ puzzles fitting in ~3MB while keeping addressable options open when puzzle quantity scale.

PLAYABLE BUILD

Try the WebGL build.

The playable build loads only when launched. On desktop it opens in a focused overlay; on mobile it opens in a new tab for a cleaner play experience.

NOTABLE IMPACT

Shipped a complete acrostic puzzle game in under a week using AI as a development accelerator.

Built the entire project, from gameplay to UI to polish, with agentic AI assistance, demonstrating how AI tools can dramatically compress development timelines without sacrificing quality.

Built a 4-stage AI pipeline that generates ~1,000 puzzles a day for pocket change.

Each stage handles a distinct part of puzzle creation, from difficulty specs to trivia generation to parameterized word pools, and the final stage has AI generate the assembly tool itself, costing roughly $1–2 per thousand puzzles.

Created cross-platform tooling that removes designers' dependency on engineering.

Puzzle designers can create, audit, and deploy content through a webapp, in-game admin mode, or Unity editor, all connected to the same pipeline via API.

Delivered a polished WebGL experience without over-engineering the architecture.

Loads in under 5 seconds on low-end devices, with 3,000+ puzzles fitting in ~3MB, and every architectural decision made with scale in mind but not prematurely optimized.

Back to Portfolio