FROM THE FIELD
Five Bugs That Passed Every Test
Operational discipline is the layer no architecture diagram shows. Five production gotchas from a multi-agent QA system, and the lazy assumption behind each one.
LOAD "*",8,1
AI Engineer · Test Automation Architect · Platform Builder
Context before LLM.
Portfolio of Dariusz Kowalski. I build multi-agent QA systems, AI pipelines and developer platforms. CDAT Pattern, Jarvis Platform, open-source distillates.
METHODOLOGY
CDAT Pattern
Components · Data · Actions · Tests. 4-layer Playwright architecture battle-tested across 9 production systems over 2 years.
ECOSYSTEMJarvis Platform
Private multi-agent QA platform. 34K LOC TypeScript, 9 microservices, 15 production pipelines. Ask for demo.
THESISContext before LLM
AI in QA does not start with "write me a test". It starts with deterministic, pre-processed context. LLM comes second.
PIPELINEWCAG Audit · 7 agents
7 agents parallel: keyboard, forms, modals, contrast, ARIA, semantic HTML, e-commerce. WCAG 2.2 AA full audit on live production.
PIPELINEPerformance Audit · 5 agents
5 agents parallel: bundle, Vue/React runtime, API calls, SSR/hydration, assets. 7h with AI vs 16h billable vs team-week classical.
PIPELINEFigma-to-Code
Deterministic Figma pipeline. CSS token mapping, pixel diff via odiff, codegen spec. Design tokens at data layer, LLM at logic only.
From the Field
View all →FROM THE FIELD
ADF Without Tears: The Full Pipeline and the Repo
The four-stage pipeline behind inline images in Jira: create, upload, resolve, embed. Plus the public AGPL repo you can clone and run against a mock Jira.
FROM THE FIELD
ADF Without Tears: The 303 Trick for Inline Images in Jira
Upload a screenshot to Jira and you get a gray External media box, not the picture. The fix is a 303 redirect and one fetch flag. Deterministic, no LLM.
Projects
All projects →AI QA Manual · Jarvis destylat
Production-ready manual QA workflow extracted from Jarvis. Context-first pipeline: Figma MCP + Jira webhook + Playwright CLI + Claude Agent SDK. Scale: 100-200 tasks in 2-3 days vs team-week classical.
Tech stack
- TypeScript
- Node.js
- Claude Agent SDK
- Playwright
- MCP
- n8n
CDAT Pattern · Playwright architecture
Components-Data-Actions-Tests - 4-layer architectural pattern for Playwright + TypeScript. Alternative to Page Object Model. Battle-tested across 9 production systems over 2 years.
Tech stack
- TypeScript
- Playwright
- Node.js
sdet-wcag-toolkit · Public WCAG 2.2 AA pipeline
Public AGPL-3.0 distillate of multi-agent WCAG audit pipeline. 5 AI specialists reading source via Read/Grep/Glob, plus static TypeScript analyzer and Playwright + axe-core dynamic testing. A-F grading. Case study in From the Field series #01.
Tech stack
- TypeScript
- Claude Agent SDK
- Playwright
- axe-core
- MCP
sdet-brain · Persistent RAG over MCP
Local-first persistent RAG for personal Markdown corpus. Qdrant + MLX + FastAPI + FastMCP 3.0. 12 MCP tools, 213 tests, source-available. Replaces copy-paste of context across Claude Desktop / Code / OpenCode chats.
Tech stack
- Python
- FastAPI
- FastMCP
- Qdrant
- MLX
- MCP
skills-radar · Lazy-loading skill discovery for Claude Code
Open-source MCP server fixing Claude Code skill bloat. Two-Tier Discovery: ~1k token mini-index always preloaded, full SKILL.md loaded on demand. 68% token reduction at 60 skills, roughly flat at 500. Hybrid retrieval (BM25 + dense), trust tiers, 100% local Apple Silicon stack via MLX (Qwen3-Embedding-8B + Qwen3-Coder-30B rewriter/reranker). No Ollama, no HTTP, no network.
Tech stack
- Python
- FastMCP
- ChromaDB
- Qdrant
- MLX
- sentence-transformers
sdet-perf-toolkit · Context-first performance audit
Performance audit for Nuxt3/Vue3. Deterministic floor: Lighthouse median-of-5 + Core Web Vitals via Playwright + trace + bundle. 5 AI specialists for what the floor does not catch. AGPL-3.0.
Tech stack
- TypeScript
- Lighthouse
- Playwright
- web-vitals