Praveen Kumar: notes from the workbench

Praveen Kumar: notes from the workbench Long-form, measured, reproducible notes on speech AI, LLM inference and training, every number from real hardware. https://inboxpraveen.github.io/ 2026-09-20T04:13:51Z Praveen Kumarhttps://inboxpraveen.github.io/ Who spoke when, on your own data: training, tuning and shipping a Kaldi speaker diarizer that runs on a CPU https://inboxpraveen.github.io/blogs/blog-5.html 2026-09-20T00:00:00Z 2026-09-20T00:00:00Z

Train Kaldi's x-vector diarizer on your own recordings and ship it as a CPU container: back-end training, VBx, speech detection, how much data you need, speed, and the Sortformer comparison.

<a href="https://inboxpraveen.github.io/blogs/blog-5.html">Read the post</a> Teaching a language model a language it never knew: continued pretraining, supervised fine-tuning and reinforcement learning on one laptop https://inboxpraveen.github.io/blogs/blog-4.html 2026-09-15T00:00:00Z 2026-09-20T00:00:00Z

An English-only 135M model taught Hindi on one 8 GB laptop GPU: tokenizer extension, continued pretraining, supervised fine-tuning and GRPO with verifiable rewards, every stage measured.

<a href="https://inboxpraveen.github.io/blogs/blog-4.html">Read the post</a> Where the bandwidth line ends: how Cerebras gets to 2,000 tokens a second, worked out from the outside https://inboxpraveen.github.io/blogs/blog-3.html 2026-09-14T00:00:00Z 2026-09-20T00:00:00Z

The bandwidth line predicts 600,000 tokens a second for four Cerebras wafers and they get 2,000. Where the rest goes, worked out from public sources with the arithmetic shown.

<a href="https://inboxpraveen.github.io/blogs/blog-3.html">Read the post</a> Before you tune anything: a field guide to inference speed and accuracy on the hardware you actually have https://inboxpraveen.github.io/blogs/blog-2.html 2026-09-12T00:00:00Z 2026-09-20T00:00:00Z

The one number that transfers between machines, what six model architectures do to it on CPU and GPU, a pre-check script, a ten-minute accuracy test and which quant to pick.

<a href="https://inboxpraveen.github.io/blogs/blog-2.html">Read the post</a> Getting the most tokens per second out of one laptop with llama.cpp https://inboxpraveen.github.io/blogs/blog-1.html 2026-09-07T00:00:00Z 2026-09-20T00:00:00Z

Every knob llama.cpp exposes, measured on one laptop with one Q8_0 model: thread pinning, CPU backends, quantization, GPU offload, KV cache types and speculative decoding, for speed and accuracy.