A gallery of applications built unattended by Cadillac — an autonomous code-building agent.
Every project in this repository was produced by Cadillac with zero human edits to the code. Each one was generated from a single-sentence task description, run through Cadillac's full pipeline (SPEC → PLAN → BUILD → 12-check VALIDATE → completeness CRITIC → runtime flow verification → PACKAGE), and shipped as-is.
This is not a curated showcase. It's the result of scanning the ~99 build workspaces accumulated over 8 weeks of Cadillac development and publishing every one that's actually finished, runnable, and not broken. 26 made the cut. The other 73 were mid-scaffold failures (mostly from earlier versions of the harness), duplicate task attempts, or builds whose entry points crashed.
Each project has its own README.md inside its directory that lists exactly which validation checks passed and which didn't. No project is dressed up to look better than it is.
| Project | Stack | Status |
|---|---|---|
| matrix-doom-fps | Python + pygame | All checks PASS — Matrix-themed ASCII Doom raycaster, 31 files |
| pygame-snake | Python + pygame | All checks PASS — classic Snake with pure-logic pytest tests |
| ascii-pong-curses | Python (stdlib) | All checks PASS — two-player curses Pong with --test mode |
| midnight-library-escape | Python | Game runs end-to-end; tests fail (text-adventure escape room, 5 puzzle rooms) |
| llm-token-stream-visualizer | Python + pygame | All checks PASS — vLLM tokens as glowing particles on neural-net graph |
| gpu-cluster-dashboard | Python + pygame | UI works; functional + tests flagged on empty module dirs (nvidia-smi + /proc dashboard) |
| aurora-ascii-art | Python (stdlib) | CLI works; naming/tests fail (util/signal.py shadows stdlib) |
| Project | Stack | Status |
|---|---|---|
| auction-house-marketplace | Flask + SQLite | All 10 checks PASS — eBay-style auction marketplace, 63 files |
| flask-habit-tracker | Flask + SQLite | All 12 checks PASS — habit tracker w/ auth, streaks, search |
| flask-crm | Flask + SQLite + vanilla JS | All 8 checks PASS — full-stack CRM with Kanban frontend |
| flask-todo-blueprints | Flask + SQLite | All 8 checks PASS — todo API with blueprints |
| pawperfect-react | React + Vite + TypeScript | Site builds + previews; vitest config issues (pet grooming marketing site) |
| multi-room-chat-server | Python + Flask | All 8 checks PASS — TCP chat server + Flask admin REST API |
| book-library-api | TypeScript + Express + zod | All 8 checks PASS — book library REST API w/ input validation |
| dungeon-master-api | TypeScript + Express | API responds to curl; jest setup issues (text dungeon-crawl REST API) |
| url-shortener-ts | TypeScript + Express | All 8 checks PASS — URL shortener w/ rate limiting + click stats |
| tasks-rest-api-ts | TypeScript + Express | All 8 checks PASS — tasks REST API w/ filter/sort/pagination |
| Project | Stack | Status |
|---|---|---|
| expense-tracker-cli | Python + SQLite | CRUD works — CLI with recurring transactions, budgets, CSV import |
| taskmgr-python | Python + argparse | All 8 checks PASS — todo CLI with JSON storage |
| click-calc-cli | Python + Click + rich | All 8 checks PASS — calculator with rich-rendered output |
| aiosqlite-bookmark-manager | Python + Click + aiosqlite | All 8 checks PASS — async-SQLite bookmark manager |
| log-analyzer-stdlib | Python (stdlib) | All 8 checks PASS — multi-format log analyzer (Apache/syslog/JSON) |
| csvstats-cli | TypeScript | CLI parses + computes correctly; vitest harness issues |
| cipher-forge | TypeScript | Round-trips work; some vitest setup issues (Caesar/Vigenère/Atbash/Rail/Morse) |
| oracle-of-winds | TypeScript | Systems produce stable output; some vitest setup issues (deterministic divination) |
| Project | Stack | Status |
|---|---|---|
| imageclassifier-cnn | Python + PyTorch | Train/predict run; one functional check fails on edge-case tensor shape |
Cadillac produces code that runs unattended on the kind of varied stack a developer encounters day-to-day. Across these 26 builds:
- Stacks span Python (stdlib, Flask, Click, pygame, curses, PyTorch, aiosqlite), TypeScript (Express, React, Vite, zod, vitest, jest), and the standard tooling around each.
- Projects span CLIs, web APIs, full-stack web apps, games, real-time visualizers, and an ML training pipeline.
- Validation coverage varies — some hit perfect green on all 12 checks; others have honest gaps that are documented inside each repo.
Cadillac is not magic. Roughly 1 in 4 attempted builds shipped cleanly. The other 73 workspaces either failed mid-scaffold, attempted duplicate tasks (where a later run produced a cleaner version), or had broken entry points. That ratio gets better with each version of the harness as resilience layers (operational gates, completeness CRITIC, runtime flow verification, surgical mode, progressive tiers) close more failure modes.
The point of this gallery isn't "every Cadillac build is perfect." It's "Cadillac builds real things that real people could pick up, run, and extend." Pick one. Clone it. Run it. See for yourself.
Each subdirectory is self-contained. To run any project:
git clone https://github.com/mtecnic/cadillac-builds.git
cd cadillac-builds/<category>/<project>
cat README.md # follow the quick-start thereThe per-project README has the exact commands that work for that project — usually pip install -r requirements.txt && python main.py --help for Python, or npm install && npm run dev for TypeScript.
Every project here was produced by Cadillac. The harness:
- Expanded the one-sentence task into 15-40 user stories (SPEC phase)
- Designed the architecture (PLAN phase — flat for small projects, modular for ≥15-file projects)
- Installed dependencies + scaffolded files (DEPS + SCAFFOLD)
- Wrote and debugged the implementation unattended (BUILD — read errors → edit → re-run → repeat)
- Ran a 12-check validation gate including operational deploy-readiness probes
- Audited completeness against the spec (CRITIC) and drove real user flows against the running artifact (RUNTIME) before declaring done
- Bounced back to BUILD if anything missed — with surgical-mode targeted fixes when the regular iterate loop got stuck
No humans edited the code. The original task descriptions are preserved in each project's README.
See mtecnic/cadillac for how the harness works, the wiki for architecture deep-dives, or mtecnic/cadillac README's "Real shipped builds" section for the project index.
Apache 2.0 across the whole repository. Each subdirectory inherits the top-level license.
🤖 Built by Cadillac. Honesty over polish; proof over claims.