runtimelab▮
Sign in
← All career paths

~/runtimelab $ cat ./roadmap/ai

AI/ML Engineer

From what a token is up to a served, retrieval-augmented, tool-using system.

3 challengesends in: Serve a retrieval-augmented, tool-using model in production.

  1. 1.Learn Python by writing it

    14 stages

    Build a concurrent service health checker with nothing but the standard library. Fourteen exercises, each a working part of the tool, each built around a mistake that bites.

    Beginnernot started

  2. 2.Build your own BPE Tokenizer

    14 stages

    Build a byte-pair-encoding tokenizer from scratch — training, encoding, decoding, special tokens, and the byte-level edge cases nobody warns you about. The GPT-2/GPT-4 tokenizer, cross-checked against tiktoken.

    Beginnernot started

  3. 3.Build your own Text Generator

    15 stages

    Build a text generator from counting alone — bigrams, sampling, temperature, top-k and top-p, then n-grams, smoothing, a repetition penalty and beam search. Every decoding knob in a modern LLM, with no neural network underneath.

    Beginnernot started

  4. 4.Build your own Embeddings

    coming soon

    Build word and sentence embeddings from scratch — co-occurrence counts, PPMI, truncated SVD, then skip-gram with negative sampling — and measure them against analogy and similarity benchmarks instead of eyeballing a nearest-neighbour list.

    Intermediate

  5. 5.Build an LLM (from scratch)

    coming soon

    Build a GPT-style language model from scratch — tokenizer, attention, training, and sampling, from raw tokens to generated text. Deterministic and verified end to end.

    Advanced

  6. 6.Build your own Semantic Search

    coming soon

    The retrieval system around the index, not the index itself — chunking, query and document asymmetry, BM25, reciprocal-rank fusion and reranking, scored with recall@k, nDCG and MRR. It calls a nearest-neighbour index; vector-db is the course that builds one.

    Advanced

  7. Destination

    Serve a retrieval-augmented, tool-using model in production.