The Dataset Viewer has been disabled on this dataset.

AI Model Field Tests in Real Products

I am a quality engineer (Lean Six Sigma Black Belt, supplier quality in the automotive industry) and I build AI products in my own time. These reports document how AI models behave inside real products with real users and real data, not on benchmarks: what the model had to do, what it delivered, where it failed, and how I worked around it.

Each report follows the same structure: What this is about · How it was tested · Example · Requirement · Result · Assessment · Feedback for the model vendor.

Reports (November 2026)

Report Models Verdict
Root Index: AI reads food labels Claude Sonnet 4.6 → 5.5, Haiku 4.5 → 5.5 Met with safeguards: fast and accurate, but only because uncertain digits may be left blank
Content Studio: AI agents for hooks and Reels Claude Sonnet/Haiku 5.5, Gemini 3.8 Flash, Gemma 4 12B · Kling 3.0 / O3, Veo 3.1 Fast Text meets requirements after prompt rework; Veo best for people, Kling unsuitable for on-screen content
Foldair: Apple's on-device model as a mail agent Apple Foundation Models (on-device, macOS 26) Met only as a helper: unreliable alone, reliable behind rules
Tanja's brain: local vs cloud Mistral-Nemo 12B, Qwen3 8B/14B (local) · Qwen3 14B, Llama 3.3 70B, gpt-oss-120b (cloud) Cloud models from ~70B meet the bar; none of the local ones do

Feedback

Questions, corrections, or a pre-release model you would like tested in real products? Open a thread in the Community tab of this page, or message me on LinkedIn. More projects: foldair.de/portfolio.

Downloads last month
6