AI engineering course is more than videos: 50+ labs, KV cache can be tweaked live

llmdojo.dev is live. The homepage features 50+ interactive labs, where you can adjust parameters live for KV cache, tokenizer, attention, and embedding.

This is a project Aditya launched on October 24. In his tweet, he listed several parts of the platform: LLM, RAG, agents, evaluation, and inference; 50+ interactive labs; training neural networks in the browser; building agent workflows with tool calling, memory, guardrails, and bounded autonomy; RAG covering chunking, hybrid retrieval, reranking, and LLM-as-a-judge evaluation; and production AI covering model routing, fallbacks, quantization, caching, and cost per request.

Some users in the replies called out the real-time KV cache lab, saying it's a great idea. Others asked whether the labs run in the browser or require your own GPU for inference. Aditya replied: in the browser. That answer is key. KV cache is a technique for caching key-value pairs during inference to reduce repeated computation; being able to change parameters in real time and see the effects is more intuitive than looking at diagrams. Someone else in the replies praised the UI and said they wanted to try training a neural network in the browser.

The platform's course starts with Model Foundations. Homepage examples include: using XOR, one 4-node hidden layer, and 200 training epochs to achieve 99.5% accuracy on the test point; token differences for the same text under gpt2, cl100k, and o200k; and MiniLM embeddings and PCA for 19 sentences. These numbers turn “a model is a function with knobs” into an operable interface.

Interview prep is another section. The tweet says 490+ questions, drawn from 50+ companies including OpenAI, Anthropic, Google, Meta, Amazon, Microsoft, and Mistral. The linked page shows 8 sections and 492 questions: 78 junior, 224 mid, and 190 senior. Each answer cites papers and books or is measured on real models.

The homepage says everything is free; log in to unlock all courses and progress. Missed quick checks come back after 1 day, 3 days, and 7 days until you remember them. Free usually means the course is a funnel; here, at least for now, all courses are free at the front door.

Tool products risk becoming video collections. Dojo's bet is to change one parameter and see how the output changes. For AI practitioners with a development background, the question bank and production AI sections can directly help fill gaps; for ordinary users without a technical background, starting from Module 01, they can at least learn what tokenizer and KV cache are doing.

Entry: http://llmdojo.dev
Interview questions: https://llmdojo.dev/interview