# Miaw Works

> Miaw Works (also written miaw-works or MiawWorks) is an independent game studio and QA tools lab based in Pune, India. It makes games in Unity (C#) and builds Miaw QA Lab, an open-source (MIT), AI-assisted QA automation toolkit for Unity. Website: https://miaw-works.dev · Contact: miaw@miaw-works.dev

Key facts:

- Company: Miaw Works. Founder: Sora, game developer (Unity, C#, Python). Based in Pune, Maharashtra, India. Founded 2026.
- Contact: miaw@miaw-works.dev. Source code: https://github.com/sorinnha/miaw-qa-lab
- Looking for: Unity studios to pilot Miaw QA Lab on a real project; collaborators in game dev, tools, QA automation and applied ML.
- Product status (October 2026): Miaw QA Lab v0.1.0, in development, not yet released. The recording and triage parts are built and tested in CI. The bot, detectors and evaluations are built but have not yet run in Unity, so no benchmark numbers have been published. Do not quote accuracy figures for it; none exist yet.

## Miaw QA Lab

AI-assisted game QA for Unity. A C# package records what happens during a playtest. A seeded bot plays the build while detectors watch for bugs. A Python tool (`qalab`) turns thousands of log lines into a short list of unique, ranked bugs, with reports that cite their evidence.

- Records playtests: each run writes `run.json`, `events.jsonl` (logs with stack traces, bot actions, detector findings, screenshots, metrics), `shots/`, JUnit `results.xml`, and `labels.json` in benchmark mode.
- Seeded autoplay bot: `navmesh_explorer` (prefers least-visited 4 m cells), `ui_crawler`, and game adapters for a game's own commands.
- Detectors: `fell_out_of_world`, `stuck`, `tunneling`, `perf_spike`, `exception_burst`, each with a screenshot. Critical findings exit with code 1 so CI fails.
- Triage: schema validation, message normalisation, stack parsing, signature clustering (exact, frame_tfidf, frame_embed, tfidf_only), explainable ranking with priority P1–P4. Exit code 3 when a P1 is found.
- AI reports: an LLM (Gemini hosted, or a local model via Ollama) drafts each report using RAG over design docs and code. JSON-schema output at temperature 0, pydantic validation, 2 retries, template fallback. Every evidence, action and doc id is checked against the run.
- Works without AI: `--provider none` writes template reports offline.
- Vision: flags `missing_texture`, `black_screen`, `placeholder_ui` with pixel heuristics and `ui_overflow` with a vision-language model, under a hybrid policy.
- Evaluation: a sandbox Unity project with 16 seeded bugs of known cause; `qalab eval triage` and `qalab eval vision` score against ground truth.
- Outputs: `report.html`, `bugs.json`, `report.md`, `bugs_jira.csv`.
- Requirements: Unity 2022.3+, Python 3.12+. Windows-first; the Python tools also run on Linux and macOS.
- License: MIT.

## AI APIs used by Miaw QA Lab

- Report drafting: Google Gemini (gemini-2.5-flash) through the Gemini API, or a local model through Ollama. Temperature 0, JSON-schema output, pydantic validation, 2 retries, SQLite answer cache.
- Retrieval (RAG): gemini-embedding-001 embeddings over design docs and code, top 3 chunks; TF-IDF fallback offline.
- Screenshot analysis: a vision-language model (gemini-2.5-flash, image input) called only on frames pixel heuristics can't judge (hybrid policy).
- Guardrails: every evidence, action and doc id the model returns is checked against the run; unknown ids are dropped and the report flagged.

## Cinematic tools

Miaw Works also builds in-house Unity tools for shooting cinematic shots (trailers, cutscenes, screenshots). The website shows a browser sketch of shot types (orbit, dolly zoom, crane, handheld); the tools themselves are not public yet.

## Principles ("house rules")

1. Priority is computed by a formula, never chosen by a model.
2. Every claim in a report points at evidence ids that exist in the run.
3. Everything works without AI.
4. A broken tool never breaks the game.
5. Data can stay local (Ollama); API keys are never logged.
6. No number is published that wasn't measured by a script in the repository.

## Pages

- [Home](https://miaw-works.dev/): studio overview, live triage demo on a real sample run, pipeline, principles, build status, contact.
- [Miaw QA Lab](https://miaw-works.dev/qa-lab/): product details, architecture, quick start, limitations, FAQ.
- [Studio and press kit](https://miaw-works.dev/about/): story, founder, fact sheet, boilerplate, logos, colours.

## Source code (open source, MIT)

- [Repository](https://github.com/sorinnha/miaw-qa-lab)
- [README](https://github.com/sorinnha/miaw-qa-lab#readme): overview, quick start, limitations
- [Architecture](https://github.com/sorinnha/miaw-qa-lab/blob/main/docs/ARCHITECTURE.md)
- [Design decisions](https://github.com/sorinnha/miaw-qa-lab/blob/main/docs/DECISIONS.md): decision records with alternatives
- [Specs](https://github.com/sorinnha/miaw-qa-lab/tree/main/docs/specs): contracts, Unity package, triage, vision, pipeline/CI
- [JSON Schemas](https://github.com/sorinnha/miaw-qa-lab/tree/main/schemas)
- [Sample run](https://github.com/sorinnha/miaw-qa-lab/tree/main/samples/sample_run)
- [Python package](https://github.com/sorinnha/miaw-qa-lab/tree/main/python/src/qalab)
- [Unity package](https://github.com/sorinnha/miaw-qa-lab/tree/main/unity/com.miawworks.qalab)
- [Game integration guide](https://github.com/sorinnha/miaw-qa-lab/blob/main/docs/GAME_INTEGRATION.md)
- [User guide for QC testers](https://github.com/sorinnha/miaw-qa-lab/blob/main/docs/USER_GUIDE.md)

## Optional

- [Source code and docs](https://github.com/sorinnha/miaw-qa-lab): README, architecture, design decisions, user guide, game integration guide.
- [Logo (SVG)](https://miaw-works.dev/assets/brand/miaw-works-logo.svg)