03writing & teaching

the blog

deep dives on machine learning, nlp, llms and operating systems — worked through by hand, with the math and the pictures.

technical blogs 20

  1. 01 system design, visually63 infographic slides covering system-design foundations, trade-offs, data, messaging, and recovery
  2. 02 week 1 — linear regressiona deep dive into linear regression and the math behind it
  3. 03 week 2 — transformersa deep dive into transformers and a visual guide into how it works
  4. 04 week 3 — optimizersa deep dive into optimizers and a journey through time
  5. 05 week 4 — rnna deep dive into recurrent neural networks and how the math behind it works
  6. 06 week 5 — basics of nlp [part 1]text preprocessing, regex, frequencies, and word embeddings
  7. 07 week 6 — basics of nlp [part 2]pos tagging, ner, sentiment analysis, and n-gram models
  8. 08 week 7 — basics of nlp [part 3]hidden markov models, text clustering, and attention
  9. 09 week 8 — llms [part 1]embeddings, positional embeddings, and tokenizers (especially bpe)
  10. 10 week 9 — enhance your model [part 1]lora, model distillation, gradient clipping and early stopping
  11. 11 week 10 — llms [part 2]how attention works, in great detail
  12. 12 deepseek r1 explanationhow deepseek r1 works under the hood
  13. 13 all about quantizationhow quantization occurs in llms
  14. 14 training translation models on low-resource languageshow to train translation models on low-resource languages
  15. 15 what is kl divergenceunderstanding kl divergence and its applications in machine learning
  16. 16 how does virtual memory even work?virtual memory, page tables, tlbs, page faults, and copy-on-write
  17. 17 how does interprocess communication work?pipes, message queues, shared memory, and synchronization
  18. 18 how does dynamic memory allocation work?the heap, malloc, free, and allocator internals
  19. 19 how do keyword spotting models work?audio features, neural networks, training, and continuous detection
  20. 20 turboquanthow polarquant and quantized johnson-lindenstrauss combine to achieve 3-bit kv-cache compression with zero memory overhead