The real job (not the LinkedIn version)You'll be working on Octo, our AI interviewer that's basically trying to understand if humans are bullshitting or not. This isn't "let's throw TensorFlow at it and hope" - this is building systems that can detect subtle patterns in how people speak, understanding context across languages, and figuring out if someone actually knows React or just memorized some Medium articles.

Day to day, you'll be:

• Building models that can understand when a candidate is explaining a concept versus just reciting definitions

• Training systems on interview data from 25+ languages without losing context

• Creating evaluation pipelines that can assess technical skills from conversational data

• Debugging why the model thinks every JavaScript developer is also a coffee expert (real bug we had)

• Building real-time inference systems that can handle video interviews without melting our servers

• Writing papers? Maybe. Shipping production code? Definitely.

You probably have:

• Built and deployed actual ML systems that real people use (not just Kaggle competitions)

• Deep understanding of transformers, embeddings, fine-tuning (you know why attention is all you need)

• Experience with multi-modal models (video + audio + text)

• Opinions about PyTorch vs TensorFlow (and can defend them)

• The ability to explain complex ML concepts to our sales team without making them cry

• A GitHub profile that shows you building weird AI stuff for fun

Epic work that would make us call you immediately:

• You've contributed to major ML frameworks

• Your model is serving millions of predictions daily somewhere

• You've published at NeurIPS, ICML, or similar (or you could have but chose to build instead)

• You built an AI product that got acquired

• You've been banned from any ML competition for being too good

Why most people fail this interview:They can't explain why their model makes specific decisions. Or they've only worked with pre-trained models. Or they think more parameters always equals better. Or they can't write production code, just notebooks.