17 Years of Experience in Kubernetes: How to Catch a Fake Resume and Avoid Spending Money on Neural Networks

A resume I reviewed not long ago claimed 17 years of Kubernetes experience. Kubernetes went public in 2014. This is not a funny outlier, it is the median of the incoming stream, and every such resume first lands on a tired recruiter who is not obligated to remember when Rust, Python 3.11 or Kubernetes shipped.

I build GetPruf, a service that screens resumes before the interview, and this talk is about the engineering problem behind it: catch the fabrications, do not go broke on the model bills, and do not hand the hiring decision to a model.

Each of the three is clear on its own, together they conflict: a flagship model on every request bankrupts you, a cheap-only model scores below a junior, and open-source through the cloud will eventually return 429 across the whole key pool at once — usually during the investor demo.

I show how to fix this with a three-tier cascade where the expensive model runs last, not first, and most of the input never reaches it. On top of the cascade there are three guard layers: handwritten deterministic rules between the model and the scoring that throw out the impossible (Kubernetes release year minus the claimed start date equals a negative number), a final output shaped as a recommendation for a human rather than a verdict from the model (which keeps it on the right side of GDPR, the EU AI Act and NYC Local Law 144), and a gateway that routes requests across AWS Bedrock, Ollama Cloud, Yandex AI Studio and the Anthropic API without a single branch in the code.

There will be concrete numbers before and after each tier, a recipe for handling 429 across several providers at once, and a collection of real fabrications this cascade caught. Anonymized, of course. But impressive.