---
title: "Sachin Rajput (@sachinrajput) | Hashnode"
description: "AI Inference & LLM Serving Engineer · latency, throughput & cost per token. Read the latest articles by Sachin Rajput on Hashnode."
url: "https://hashnode.com/@sachinrajput"
---

@sachinrajput

# Sachin Rajput

@sachinrajput · Gurugram, India · Joined October 2026

AI Inference & LLM Serving Engineer · latency, throughput & cost per token

[Website](https://sachinrajput.dev/)

### [The KV Cache From Scratch](https://sachinrajput.hashnode.dev/the-kv-cache-from-scratch?utm_source=hashnode&utm_medium=feed)

We build the model's notebook by hand, prove it gives the same answer 15 times faster, and weigh it in bytes.

10 min read · [sachinrajput.hashnode.dev](https://sachinrajput.hashnode.dev/the-kv-cache-from-scratch?utm_source=hashnode&utm_medium=feed)

[Nov 4](https://hashnode.com/posts/the-kv-cache-from-scratch/6aeb0e096046a33c40343a36)

### [The Roofline Model for LLM Inference](https://sachinrajput.hashnode.dev/the-roofline-model-for-llm-inference?utm_source=hashnode&utm_medium=feed)

Two speed limits, one picture. Measured on a real computer, then used to predict how fast a GPU can type.

12 min read · [sachinrajput.hashnode.dev](https://sachinrajput.hashnode.dev/the-roofline-model-for-llm-inference?utm_source=hashnode&utm_medium=feed)

[Nov 2](https://hashnode.com/posts/the-roofline-model-for-llm-inference/6ae8816cccbfbf7d7a241be8)

### [Prefill vs Decode: Two Phases, Two Bottlenecks](https://sachinrajput.hashnode.dev/prefill-vs-decode-two-phases-two-bottlenecks?utm_source=hashnode&utm_medium=feed)

Reading your question gets stuck on math. Writing the answer gets stuck on memory. Measured, then mixed on one server.

12 min read · [sachinrajput.hashnode.dev](https://sachinrajput.hashnode.dev/prefill-vs-decode-two-phases-two-bottlenecks?utm_source=hashnode&utm_medium=feed)

[Oct 29](https://hashnode.com/posts/prefill-vs-decode-two-phases-two-bottlenecks/6ae3109fa96cdd7360256afa)

### [A Forward Pass for One Token: Where the FLOPs Go](https://sachinrajput.hashnode.dev/a-forward-pass-for-one-token-where-the-flops-go?utm_source=hashnode&utm_medium=feed)

Ride along with one word as it climbs through a language model, and count every bit of math it causes.

15 min read · [sachinrajput.hashnode.dev](https://sachinrajput.hashnode.dev/a-forward-pass-for-one-token-where-the-flops-go?utm_source=hashnode&utm_medium=feed)

[Oct 26](https://hashnode.com/posts/a-forward-pass-for-one-token-where-the-flops-go/6adf26387581b6231bccba15)

### [One Request, End to End: Tokenize, Prefill, Decode, Stream](https://sachinrajput.hashnode.dev/one-request-end-to-end-tokenize-prefill-decode-stream?utm_source=hashnode&utm_medium=feed)

What really happens in the second after you ask a chatbot a question? A trip through the Word Factory, explained simply.

16 min read · [sachinrajput.hashnode.dev](https://sachinrajput.hashnode.dev/one-request-end-to-end-tokenize-prefill-decode-stream?utm_source=hashnode&utm_medium=feed)

[Oct 22](https://hashnode.com/posts/one-request-end-to-end-tokenize-prefill-decode-stream/6ad9e47163fdbe14d18a20a6)