Skip to content
LeetLLM
Tracks
Beginner23 lessons

AI Engineer Foundations

Curious about how models work? Start with small experiments and build the foundations for your first LLM app.

Software engineers, students, and career switchers new to ML systems.

You can read model diagrams, run Python experiments, call LLM APIs, and explain the first full AI app path.

  1. 1Git, Shell, Linux for AIMake a teammate's clone reprint the 0.667 eval instead of crashing on a file that stayed on your laptop.Computing FoundationsEasy21 min
  2. 2Docker for Reproducible AIPackage the access-request scorer so a fresh container can reprint 0.667 without your laptop's Python.Computing FoundationsEasy21 min
  3. 3Python for AI EngineeringRebuild the Docker lesson's 2/3 receipt in Python, and keep a bad JSONL field from looking like a wrong answer.Computing FoundationsEasy20 min
  4. 4NumPy and Tensor ShapesTwo log snippets hold 24 numbers. Naming each axis tells you which values belong to which word before an index, a broadcast, or a reduction can be trusted.Computing FoundationsEasy22 min
  5. 5Data Structures for AIThree notes can be scanned with a list. The same search gets a different cost once each operation picks a list, index, heap, queue, or cache.Computing FoundationsEasy24 min
  6. 6SQL and Data ModelingWhen the support retriever restarts and its document dictionary is empty, SQL tables keep the versioned policy, permissions, and citations that make the answer recoverable.Computing FoundationsEasy27 min
  7. 7Dataset Pipelines and Data QualityTwo incident rows that differ only by spaces can train and test on the same conversation. Gates, grouped splits, and a manifest make that leak visible.ML Algorithms & EvaluationMedium30 min
  8. 8Gradients and BackpropTrain a one-weight latency predictor by reading a slope, taking a downhill step, tracing chain-rule paths, and checking the same gradient in PyTorch.Math & StatisticsEasy25 min
  9. 9Vectors, Matrices & TensorsOne latency row is two numbers. The same multiply has to survive a batch of requests and a tensor of tokens without mixing the axes.Math & StatisticsEasy19 min
  10. 10Probability for Machine LearningA detector that catches 95 percent of abusive signups can still fill the review queue with clean accounts, and the flagged-pile counts show why.Math & StatisticsEasy23 min
  11. 11Statistics and UncertaintySixteen confirmed abuse cases in 100 random flags are not yet the queue's rate, and a tighter interval can still describe the wrong pile.Math & StatisticsEasy21 min
  12. 12Neural Networks from ScratchA timeout-risk score starts as a weighted sum. See why a non-linearity, a shape contract, and backpropagation turn that sum into a network you can train.Preparation & PrerequisitesEasy26 min
  13. 13Training & BackpropagationA latency prediction is four milliseconds low. Follow that error through one update, then see how batch size, depth, and initialization change the step.Preparation & PrerequisitesEasy28 min
  14. 14PyTorch Training LoopsA falling loss can hide a broken ticket-router loop, so follow one batch through logits, gradients, validation, and a checkpoint that still matches after reload.ML Algorithms & EvaluationMedium36 min
  15. 15The Transformer Architecture End-to-EndFollow `function returns the` through a causal decoder until the last row becomes a score for the next token.Preparation & PrerequisitesEasy26 min
  16. 16Language Modeling & Next TokensThe prompt ends with 'the'. Follow that prefix from a counted bigram through teacher-forced training to the cache that serves the next token.Preparation & PrerequisitesEasy30 min
  17. 17From GPT to Modern LLMsOne model is asked to label a review, pull RUN-842 from a log, and explain the failure. Each later design choice still needs its own check.Preparation & PrerequisitesEasy33 min
  18. 18BPE, WordPiece, and SentencePieceSee why a word like unexpectedly should split into reusable pieces, and why a new tokenizer can still point an old checkpoint at the wrong vector.Core LLM FoundationsMedium33 min
  19. 19Static to Contextual EmbeddingsThe same token ID starts from one table row, while invoice and phone clues can move each occurrence toward a different meaning.Core LLM FoundationsMedium26 min
  20. 20Prompt Engineering FundamentalsBuild an incident-triage prompt with clear roles, useful examples, structured outputs, evidence checks, and a separate authorization boundary.Preparation & PrerequisitesEasy16 min
  21. 21Calling LLM APIs in ProductionA notebook answer can be right and still stall, time out, or return HTTP 429. The server has to decide what to retry and what the user may see.Preparation & PrerequisitesEasy30 min
  22. 22First AI App End-to-EndThe user clicks Check policy and reloads. Follow whether the request finished, which account facts it used, and what the page may claim.Preparation & PrerequisitesEasy15 min
  23. 23LLM Benchmarks & LimitationsA fluent CodeAssist answer invents on-call deploy permission the freeze policy never granted. A public score will not catch that.Core LLM FoundationsMedium35 min