Turbowork

We build AI tools through a pipeline that does not trust the AI.

Turbowork is a one-founder software company in South Korea. Every product — desktop tools for sellers and creators, games, contest entries — is written by AI coding agents inside the Red-Sandwich Harness, where locked specs, hard code limits and an independent red-team auditor decide what ships.

Red-Sandwich Harness

Our core method, published as an open specification on GitHub. AI-generated code can pass a happy-path example while missing edge cases or drifting from the agreed spec. The harness wraps every build step between two deterministic gates and lets an adversarial reviewer attack it.

Top breadSpec-Lock TDD GateInput/output contracts are fixed before implementation; execution is blocked when they fail.
FillingAST Pure Units≤ 35 logic lines per function, ≤ 200 lines per file, cyclomatic complexity ≤ 5 — checked on every edit.
Bottom breadStrict IsolationFresh state per fold and per step; only the exact tree that passed audit is committed.
Red-Team Auditor

An independent model attacks the implementation with concrete counterexamples.

Missing values, extreme inputs, shape mismatches and leakage paths. Each finding is reproduced locally, then locked in as a test.

Only an explicit PASS from the executable gates admits the next stage — never the coding agent's own report.

Grown into the v6.6 deterministic harness

The sandwich now runs unattended across all our projects: one step per AI session, two separate audits, automatic retries.

  1. Spec & plan map
  2. Build (Claude Code)
  3. Lint gate (AST hooks)
  4. Rule audit (GPT)
  5. Final audit (GPT)
  6. Commit audited tree

FAIL → a new session with raised effort reads the failure rack and patches; 4th failure stops for a human decision.

97machine-checked rules (R01–R101)
110 / 110harness self-tests passed (deliberately broken dummy projects)
0leftover temp files after 900 atomic state writes under load
2independent audits per step — the builder never grades itself

Tech stack

Chosen per product; most tools run fully on the user's PC with no file uploads.

AI orchestration

Claude Code (hooks, MCP, sub-agents), OpenAI & Gemini APIs, PowerShell runner, Python tooling, memory & CPU guards

Languages

Python, TypeScript, Kotlin, Rust, C#, PowerShell

Desktop apps

Tauri 2 + React, PySide6 / PyQt6, pywebview, PyInstaller, signed release ZIPs with SHA-256

AI / media

PyTorch + CUDA, faster-whisper, TTS, Shap-E, TripoSG, FFmpeg (NVENC), Blender 5 add-ons, Remotion

Web & backend

FastAPI, Node.js, Vite, Next.js, Cloudflare Workers + D1, SQLite (WAL, FTS5)

3D & games

three.js, Rapier, cannon-es, Unity 6 (C#), Android / Android Automotive OS

Portfolio

All numbers come from each project's own run logs; nothing without a record is claimed.

Contests & hackathons

GMP AHU AI self-diagnosis

SK AI Hackathon (AI Idea League) · submitted

Diagnoses clean-room air-handling units from PLC history and SOPs; read-only, never writes to the PLC. FastAPI, SQLite, OpenAI.

4 AHU units · 7 synthetic fault scenarios

Pleos Concierge

Hyundai Motor Group in-vehicle app challenge · submitted Nov 2026

Android Automotive concierge with a pure-Kotlin safety core that blocks interaction while driving. Demo build.

core 23/23 · DEMO 48/48 tests · signed APK · GitHub · demo video

Name That LLM

Kaggle competition · in progress

Identifying which language model wrote a text. Every experiment fixes its acceptance rule before seeing results.

best public score 0.693 (Nov 15, 2026)
Products (Turbowork brand)

Vocal Studio

Tauri 2 + React, Python · on sale (Kmong)

Character TTS and narration to auto-subtitled, synced, mastered video.

84 engine tests passed

Subtitle Generator & Batch

Python, Tauri 2, local speech recognition

Korean video to SRT drafts on the user's PC, with a crash-safe batch queue.

409 regression tests · UI 22/22

Sync Subtitle Maker

Tauri 2, TypeScript, Rust · on sale

Audio + script to timed SRT using silence analysis — no AI API.

87 tests passed

CutCatch

Python, FFmpeg (NVENC) · installer 0.9.4

Screen recording and capture for Windows, with an automated lecture-recording mode.

152 tests passed

QuickPrompt CS

Python, PyQt6 · portable v1.1.1

One hotkey shows reply snippets beside the cursor and pastes into the original field.

87 tests, OK

Seller automation

Python, Naver Commerce API

Margin calculator, listing via API, order sheets for a live cross-border store; NStorePriceGuard watches competitor prices (changes locked).

528 · 172 tests passed

Dealbite

Live at dealbite.kr

Free shopping calculators today; next, a price-history deal service that filters fake discounts.

Engines, AI tools & experiments

Blender AI 3D Asset Generator

PyTorch + CUDA, Shap-E, FastAPI, Blender 5.2, C#

Text or image to PBR-textured GLB, imported straight into Blender.

263 tests · text-to-3D 38.4 s on RTX 3060

assetpipe

Python, Blender, TripoSG, Unity 6

Game asset pipeline with cost, license and SHA-256 recorded per asset.

14/14 steps · 88 tests + 8 real Unity tests

Control Center

Python, pystray, pywebview, MCP

Tray app watching PCs, servers, pipelines and MCP plugs; exposes a read-only MCP server.

760 tests passed

AI Meeting Room

Python, edge-tts, faster-whisper

Claude and Codex debate a decision in a CLI meeting with voice.

685 tests passed

Local Coder

Ollama, Aider, pywebview

Offline coding assistant on a 12 GB GPU.

147 tests passed

Taste Meter

FastAPI, Playwright

Web app that measures taste sensitivity against standard products and compares friends.

1,307 tests passed
Games & media

Stone Stacking

three.js, cannon-es, WebSocket

Calm multiplayer stone-stacking game.

release ZIP · 26 tests passed

Unboxing House

Android

Casual unboxing game; release build (AAB) ready, Google Play setup in progress.

Snowland Sandbox

Vite, TypeScript, three.js, Rapier

Fixed 60 Hz physics decoupled from display rate, capped catch-up steps.

Trot Joseon TV

YouTube channel

Korean trot sung by our original virtual singer — lip-sync, rendering and upscaling. Watch

Test counts are as recorded in each project's logs at the time noted there; some predate later changes.

Where we hit the wall

Everything above was built and run on one desktop: RTX 3060 (12 GB), i5-12400F, 32 GB RAM. Code is fine; graphics are not. A lip-sync model alone needs ~11 GB, so heavy jobs queue one at a time through our own memory guard. With cloud GPUs and credits we would render full-length virtual-singer videos, batch seller videos, and generate 3D assets at full quality.