You know how to train a model. It's everything around it that keeps breaking: the data is messier than the paper assumed, the fine-tune scores well and behaves oddly in production, and nobody can tell you whether your eval predicts anything a user actually cares about.

That's the work we'd do together.

We start with whatever you're already working on, not a syllabus: your dataset, your training script, your eval, your traces. More often than not, the thing blocking you turns out to be the data or the eval rather than the model. You'll come away with something you can run that week, and we check it against results.

In practice it tends to look like this: building an LLM-as-a-judge that agrees with your human labels; error analysis on real traces to find where a model quietly fails; choosing between SFT, LoRA, DPO and GRPO for your problem instead of whatever's trending; getting a fine-tuning run to reproduce; or reading a paper together and working out whether it applies to you at all.

If you're coming from software engineering with no ML background, we do the same thing from a different starting point. You get an order to learn things in, one project at a time, with the maths brought in when you need it and not before. I taught myself this way (fast.ai, CS231n, then Stanford CS336), so I have a decent sense of what you can safely skip for now and what will quietly cost you six months.

ML interview preparation is a separate path, and we can spend the whole time on that if you'd prefer. I've run about ten real ML hiring interviews at a startup, plus a fair number of mock ML system design rounds.

Briefly, where this comes from: 13 years in engineering, the last 8 in AI/ML. Most recently Partner Engineer for GenAI at Meta, maintaining Llama-Cookbook (18k+ GitHub stars). Before that, six years as the first engineer at an ML startup, owning data, training, evaluation and deployment end to end. I contributed the image grounding example to the Building with Llama 4 course on DeepLearning.AI.

Start the 7-day trial and send over one thing you're stuck on. We'll spend the first call on that, and you'll know inside the week whether it's worth carrying on.