<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Sreayan De</title>
    <link>https://sreayan.dev</link>
    <description>Writing about systems programming, distributed systems, and things I'm building.</description>
    <language>en-us</language>
    <atom:link href="https://sreayan.dev/feed.xml" rel="self" type="application/rss+xml"/>
    <item>
      <title>url shortening without the database</title>
      <link>https://sreayan.dev/blog/url-shortening-without-the-database</link>
      <guid>https://sreayan.dev/blog/url-shortening-without-the-database</guid>
      <pubDate>Thu, 09 Jul 2026 00:00:00 GMT</pubDate>
      <description>how a bloom filter, redis cache, and celery task queue keep redirect latencies low when sqlite gets overwhelmed.</description>
    </item>
    <item>
      <title>building a tokenizer cache from scratch</title>
      <link>https://sreayan.dev/blog/building-a-tokenizer-cache-from-scratch</link>
      <guid>https://sreayan.dev/blog/building-a-tokenizer-cache-from-scratch</guid>
      <pubDate>Wed, 01 Jul 2026 00:00:00 GMT</pubDate>
      <description>a dict, a linked list, and a lock walk into an LLM pipeline. the result is a 10x speedup on repeated tokenization.</description>
    </item>
    <item>
      <title>flashattention from first principles</title>
      <link>https://sreayan.dev/blog/flashattention-from-first-principles</link>
      <guid>https://sreayan.dev/blog/flashattention-from-first-principles</guid>
      <pubDate>Wed, 24 Jun 2026 00:00:00 GMT</pubDate>
      <description>i knew how attention worked in C++ but couldn't read the FlashAttention paper. the missing piece was one sentence: a sentence is a matrix. once that clicked, i wrote both kernels in CUDA and watched the tiled version pull 3x on a T4.</description>
    </item>
    <item>
      <title>attention is all you need, in raw c++</title>
      <link>https://sreayan.dev/blog/attention-is-all-you-need-in-raw-cpp</link>
      <guid>https://sreayan.dev/blog/attention-is-all-you-need-in-raw-cpp</guid>
      <pubDate>Wed, 17 Jun 2026 00:00:00 GMT</pubDate>
      <description>i implemented a transformer encoder from scratch in c++ with zero dependencies. matrix math, multi-head attention, SIMD matmul, 4-bit quantization, and a full forward pass you can hold in one file.</description>
    </item>
    <item>
      <title>building a distributed rate-limited http server in go</title>
      <link>https://sreayan.dev/blog/building-a-distributed-rate-limited-http-server-in-go</link>
      <guid>https://sreayan.dev/blog/building-a-distributed-rate-limited-http-server-in-go</guid>
      <pubDate>Thu, 11 Jun 2026 00:00:00 GMT</pubDate>
      <description>i built a go http server with distributed rate limiting via redis and lua, context timeouts, prometheus metrics, graceful shutdown, and load tested it at ~7,600 rps.</description>
    </item>
    <item>
      <title>redis token bucket rate limiting in go</title>
      <link>https://sreayan.dev/blog/redis-token-bucket-rate-limiter-in-go</link>
      <guid>https://sreayan.dev/blog/redis-token-bucket-rate-limiter-in-go</guid>
      <pubDate>Sun, 26 Apr 2026 00:00:00 GMT</pubDate>
      <description>i needed a real per-ip limiter across multiple go instances, so i pushed the bucket into redis and made refill + consume atomic with lua.</description>
    </item>
    <item>
      <title>the cache is lying to you</title>
      <link>https://sreayan.dev/blog/the-cache-is-lying-to-you</link>
      <guid>https://sreayan.dev/blog/the-cache-is-lying-to-you</guid>
      <pubDate>Tue, 21 Apr 2026 00:00:00 GMT</pubDate>
      <description>why the most common performance optimization in distributed systems is also one of its most dangerous failure modes</description>
    </item>
    <item>
      <title>understanding llm inference engineering: why your laptop can't keep up</title>
      <link>https://sreayan.dev/blog/understanding-llm-inference-engineering-why-your-laptop-cant-keep-up</link>
      <guid>https://sreayan.dev/blog/understanding-llm-inference-engineering-why-your-laptop-cant-keep-up</guid>
      <pubDate>Sat, 18 Apr 2026 00:00:00 GMT</pubDate>
      <description>i benchmarked a small llm on my cpu and finally saw where inference time actually goes.</description>
    </item>
    <item>
      <title>how llms actually run at scale</title>
      <link>https://sreayan.dev/blog/how-llms-actually-run-at-scale</link>
      <guid>https://sreayan.dev/blog/how-llms-actually-run-at-scale</guid>
      <pubDate>Thu, 16 Apr 2026 00:00:00 GMT</pubDate>
      <description>a systems deep dive into what it takes to serve a 405B parameter model to thousands of users.</description>
    </item>
    <item>
      <title>implementing mapreduce in go</title>
      <link>https://sreayan.dev/blog/implementing-mapreduce-in-go</link>
      <guid>https://sreayan.dev/blog/implementing-mapreduce-in-go</guid>
      <pubDate>Tue, 14 Apr 2026 00:00:00 GMT</pubDate>
      <description>building a concurrent mapreduce engine from scratch to count discord messages.</description>
    </item>
    <item>
      <title>building mini-git</title>
      <link>https://sreayan.dev/blog/building-mini-git</link>
      <guid>https://sreayan.dev/blog/building-mini-git</guid>
      <pubDate>Mon, 13 Apr 2026 00:00:00 GMT</pubDate>
      <description>i'm building git from scratch in c++.</description>
    </item>
    <item>
      <title>how i accidentally derived mapreduce</title>
      <link>https://sreayan.dev/blog/how-i-accidentally-derived-mapreduce</link>
      <guid>https://sreayan.dev/blog/how-i-accidentally-derived-mapreduce</guid>
      <pubDate>Mon, 13 Apr 2026 00:00:00 GMT</pubDate>
      <description>i tried to count discord messages and ended up reinventing a 2004 google paper.</description>
    </item>
    <item>
      <title>hello, world</title>
      <link>https://sreayan.dev/blog/hello-world</link>
      <guid>https://sreayan.dev/blog/hello-world</guid>
      <pubDate>Sun, 12 Apr 2026 00:00:00 GMT</pubDate>
      <description>first post. why i wanted a blog and why twitter is too short.</description>
    </item>
  </channel>
</rss>
