<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Manav Shah</title>
    <description>The latest articles on DEV Community by Manav Shah (@manavshaha).</description>
    <link>https://dev.to/manavshaha</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4202709%2F1e2b00f9-f3e7-41ee-a511-6325b731da71.png</url>
      <title>DEV Community: Manav Shah</title>
      <link>https://dev.to/manavshaha</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/manavshaha"/>
    <language>en</language>
    <item>
      <title>LLM API Cost Optimization: A Practical Guide With Real Numbers From a Live System</title>
      <dc:creator>Manav Shah</dc:creator>
      <pubDate>Thu, 29 Oct 2026 14:06:17 +0000</pubDate>
      <link>https://dev.to/manavshaha/llm-api-cost-optimization-a-practical-guide-with-real-numbers-from-a-live-system-4jdb</link>
      <guid>https://dev.to/manavshaha/llm-api-cost-optimization-a-practical-guide-with-real-numbers-from-a-live-system-4jdb</guid>
      <description>&lt;p&gt;Every team running LLM features eventually hits the same wall: the API bill grows faster than the product. The usual advice is a list of techniques with big percentage claims and no proof. This guide is different — every number from our own system below is measured, not estimated.&lt;/p&gt;

&lt;p&gt;I build RRC Gateway, an AI API cost reducer. We ran a live benchmark on our production proxy: 50 real requests across exact repeats, reworded questions, and fresh unique prompts. Here is what actually works, in order of impact.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Semantic caching — the biggest lever for repeat traffic
&lt;/h2&gt;

&lt;p&gt;Cache answers by meaning, not exact text. "What is your refund policy?" and "Tell me about your refund policy" are the same question to a user, but two full-price API calls to a provider. Semantic caching embeds the incoming query, matches it against cached answers above a similarity threshold, and serves the cached response without any provider call.&lt;/p&gt;

&lt;p&gt;Measured on our system: 10 out of 10 exact repeats served at $0.00. 3 out of 10 reworded questions caught by semantic caching — 29.4% savings on that slice. Support bots, internal tools, and eval loops are where this wins hardest, because their traffic repeats by nature.&lt;/p&gt;

&lt;p&gt;The honest limit: a question asked for the first time saves nothing. Fresh unique prompts in our benchmark saved 0%, exactly as expected.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Smart routing — the right model for each job
&lt;/h2&gt;

&lt;p&gt;Not every request needs a frontier model. Classification, formatting, extraction, and FAQ answering run identically on models a fraction of the price. Smart routing scores each prompt for complexity and sends simple work to cheap models automatically, reserving premium models for genuinely hard reasoning.&lt;/p&gt;

&lt;p&gt;This is where most teams overpay without noticing: one premium model handling everything, including trivial work.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Context trimming — stop resending what the model already knows
&lt;/h2&gt;

&lt;p&gt;System prompts plus retrieved documents mean every request carries far more tokens than the model needs. Retrieve only the tokens the model actually needs for this request instead of resending your whole knowledge base. In RAG setups this alone removes a large share of input tokens.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Prompt caching — for stable prefixes
&lt;/h2&gt;

&lt;p&gt;When many requests share a system prompt or context prefix, provider-side prompt caching charges a fraction for the cached portion. Structure prompts with a stable prefix and a variable suffix so the repeated part is cached, not rebilled.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Model selection and batching — the supporting levers
&lt;/h2&gt;

&lt;p&gt;Start with the cheapest model that could plausibly work, and only upgrade when you see concrete quality issues. For work nobody is waiting on — overnight tagging, evals, reports — batch APIs typically charge half the normal price.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the full picture looks like
&lt;/h2&gt;

&lt;p&gt;Across our whole 50-request mixed workload — repeats, rewordings, and fresh prompts together — we measured 26% savings, 1,826 tokens saved. That is the honest headline: your savings scale with your repeat rate. Support-style traffic with heavy repeats saves far more; a workload where every prompt is brand new saves less. Any guide promising one fixed percentage for every workload is selling, not measuring.&lt;/p&gt;

&lt;h2&gt;
  
  
  Doing it without rewriting your app
&lt;/h2&gt;

&lt;p&gt;RRC Gateway does caching, routing, and context trimming behind one OpenAI-compatible URL — change the base URL and you are done:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;OpenAI&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;YOUR_KEY&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://gateway.rrcsolutions.me/v1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It is in free public beta: gateway.rrcsolutions.me. Point it at a workload and watch your own cache hit rate — that number is the real answer to what you will save.&lt;/p&gt;

&lt;p&gt;Beyond repeat-prompt caching, in pipeline: auto training loop, auto understanding, and auto DB management.&lt;/p&gt;

&lt;p&gt;Full guide on our site: &lt;a href="https://gateway.rrcsolutions.me/llm-api-cost-optimization" rel="noopener noreferrer"&gt;https://gateway.rrcsolutions.me/llm-api-cost-optimization&lt;/a&gt;&lt;/p&gt;

</description>
      <category>llm</category>
      <category>ai</category>
      <category>api</category>
      <category>optimization</category>
    </item>
    <item>
      <title>How Indian Startups Can Cut Their OpenAI API Bill — Real Numbers, No Theory</title>
      <dc:creator>Manav Shah</dc:creator>
      <pubDate>Sat, 24 Oct 2026 13:24:00 +0000</pubDate>
      <link>https://dev.to/manavshaha/how-indian-startups-can-cut-their-openai-api-bill-real-numbers-no-theory-2ijl</link>
      <guid>https://dev.to/manavshaha/how-indian-startups-can-cut-their-openai-api-bill-real-numbers-no-theory-2ijl</guid>
      <description>&lt;p&gt;If you are building an AI startup in India, your OpenAI bill hurts twice: once in dollars, once in rupees. Every wasted token is billed in USD and felt in INR.&lt;/p&gt;

&lt;p&gt;I build RRC Gateway, an AI API cost reducer. Last week I ran a live benchmark on our own production proxy — 50 real requests, measured, not estimated. Here is what it showed, and what it means in rupees.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Indian startups overpay
&lt;/h2&gt;

&lt;p&gt;Most Indian AI startups I talk to have the same three leaks:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Repeat questions.&lt;/strong&gt; Support bots, internal tools, eval loops — the same questions asked slightly differently, answered from scratch every time. "What is your refund policy?" and "Tell me about your refund policy" are two full-price API calls.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bloated context.&lt;/strong&gt; System prompts plus retrieved docs mean every request carries far more tokens than the model needs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Premium models doing trivial work.&lt;/strong&gt; GPT-4o classifying sentiment, formatting JSON, answering FAQs — work a model 10x cheaper could do identically.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Real numbers from a live benchmark
&lt;/h2&gt;

&lt;p&gt;The benchmark: 50 requests through the gateway, three slices — exact repeats, reworded questions, and fresh unique prompts.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Exact repeats: 10 out of 10 served from cache at $0.00.&lt;/strong&gt; Every single one. 100% savings on those calls.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reworded questions: 3 out of 10 caught by semantic caching.&lt;/strong&gt; 29.4% savings there.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Whole mixed workload: 26% savings.&lt;/strong&gt; 1,826 tokens saved across 50 requests.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;No cherry-picking: fresh unique prompts saved 0%, exactly as expected — there is nothing to cache the first time. The honest headline is this: &lt;strong&gt;your savings scale with your repeat rate.&lt;/strong&gt; Support-style traffic with heavy repeats saves far more than a workload where every prompt is brand new.&lt;/p&gt;

&lt;h2&gt;
  
  
  What 26% means in rupees
&lt;/h2&gt;

&lt;p&gt;Illustrative math, clearly labelled: if your startup spends $1,000 a month on OpenAI, 26% is $260 a month — roughly ₹23,000 a month at ~₹88 to the dollar, or about ₹2.7 lakh a year. That is a hire, a marketing budget, or several more months of runway. And that is the &lt;em&gt;mixed&lt;/em&gt; number — if your traffic is support-heavy, the ceiling is far higher.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to reduce OpenAI API costs in India
&lt;/h2&gt;

&lt;p&gt;The levers, in order of impact:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Semantic caching.&lt;/strong&gt; Cache answers by meaning, not exact text, so reworded duplicates never reach the provider.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Smart routing.&lt;/strong&gt; Score each prompt for complexity; send simple work to cheap models automatically.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Context trimming.&lt;/strong&gt; Retrieve only the tokens the model actually needs instead of resending your whole knowledge base.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;RRC Gateway does all three behind one OpenAI-compatible URL — change the base URL and you are done:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;OpenAI&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;&amp;lt;redacted&amp;gt;&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://gateway.rrcsolutions.me/v1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It is in free public beta: gateway.rrcsolutions.me. If you point it at a workload, I genuinely want to hear your cache hit rate — that is the number I obsess over.&lt;/p&gt;

</description>
      <category>openai</category>
      <category>llm</category>
      <category>ai</category>
      <category>india</category>
    </item>
    <item>
      <title>I Kept Paying for the Same AI Answers — So I Built a Gateway That Caches Them</title>
      <dc:creator>Manav Shah</dc:creator>
      <pubDate>Fri, 23 Oct 2026 13:30:16 +0000</pubDate>
      <link>https://dev.to/manavshaha/i-kept-paying-for-the-same-ai-answers-so-i-built-a-gateway-that-caches-them-3nf8</link>
      <guid>https://dev.to/manavshaha/i-kept-paying-for-the-same-ai-answers-so-i-built-a-gateway-that-caches-them-3nf8</guid>
      <description>&lt;h1&gt;
  
  
  I Kept Paying for the Same AI Answers — So I Built a Gateway That Caches Them
&lt;/h1&gt;

&lt;p&gt;Last month I looked at our OpenAI bill and felt sick. Not because the number was huge — because I recognized the &lt;em&gt;questions&lt;/em&gt;. We were paying full price, every single day, for answers the model had already given us.&lt;/p&gt;

&lt;p&gt;That was the moment &lt;a href="https://gateway.rrcsolutions.me" rel="noopener noreferrer"&gt;RRC Gateway&lt;/a&gt; started.&lt;/p&gt;

&lt;h2&gt;
  
  
  The three leaks in every LLM bill
&lt;/h2&gt;

&lt;p&gt;After digging through our usage, nearly all the waste came from three places:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Duplicate prompts.&lt;/strong&gt; Support bot, internal tools, eval loops — the same questions phrased slightly differently, answered from scratch every time. "What's your refund policy?" and "Tell me about your refund policy" are two full-price API calls.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Bloated context.&lt;/strong&gt; Our system prompt plus retrieved docs meant every request carried ~100K tokens of context. The model needed maybe 2,000 of them. We paid for all 100K, every call.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Premium models doing trivial work.&lt;/strong&gt; GPT-4o was classifying sentiment, formatting JSON, answering FAQs. Work a model 10x cheaper could do identically.&lt;/p&gt;

&lt;p&gt;Individually these look small. Multiplied across thousands of requests a day, they &lt;em&gt;are&lt;/em&gt; the bill.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why I didn't just use the existing fixes
&lt;/h2&gt;

&lt;p&gt;Prompt compression (LLMLingua-style) shrinks tokens but you still pay per call. Manual model routing works until someone forgets and hardcodes the expensive model again. Exact-match caching misses every rephrased duplicate. Each fix solves one leak. I wanted one layer that solved all three, without touching application code.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I built: a drop-in proxy with three layers
&lt;/h2&gt;

&lt;p&gt;RRC Gateway sits between your app and the providers (OpenAI, Claude, Gemini, Mistral, Llama). You change three lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;OpenAI&lt;/span&gt;

&lt;span class="c1"&gt;# Before
&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;sk-...&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="c1"&gt;# After — everything else stays identical
&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;rrc_your_key&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://gateway.rrcsolutions.me/v1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="n"&gt;response&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;gpt-4o&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;What&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;s your refund policy?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}],&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No SDK changes, no rewrites. Behind that URL, three things happen per request:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Semantic caching.&lt;/strong&gt; The gateway embeds each incoming prompt and compares it against cached ones. "What's your refund policy?" and "Tell me about your refund policy" match at high similarity — the second one never reaches the provider. In live testing, a repeat prompt came back as a cache hit: 116 tokens served, &lt;strong&gt;$0.00 charged&lt;/strong&gt;. A cache hit is the cheapest API call there is — the one you never make.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Virtual memory.&lt;/strong&gt; Instead of resending your whole knowledge base every call, your docs are compressed into vector embeddings once. Each request retrieves only the ~2,000 tokens the model actually needs. Same answers, roughly 95% fewer context tokens billed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Smart routing.&lt;/strong&gt; Every prompt is scored for complexity in real time. Genuinely hard reasoning goes to the frontier model; the FAQ, the classifier, the formatter go to a cheap one automatically. Nobody has to remember to pick — the default stops being the most expensive option.&lt;/p&gt;

&lt;h2&gt;
  
  
  The math
&lt;/h2&gt;

&lt;p&gt;Take a workload doing 50,000 requests/day against GPT-4o with heavy context. Back-of-envelope, illustrative — your numbers will differ:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Leak&lt;/th&gt;
&lt;th&gt;Typical saving on the portion it touches&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Semantic cache hits (30–50% of support-style traffic)&lt;/td&gt;
&lt;td&gt;100% of those calls&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Context trimmed 100K → 2K tokens&lt;/td&gt;
&lt;td&gt;~95% of input cost&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Simple tasks routed to cheap models&lt;/td&gt;
&lt;td&gt;10x per routed call&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Stack them and "up to 90% off the total bill" stops sounding like marketing. It sounds like arithmetic.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;RRC Gateway is in free public beta — no credit card, OpenAI-compatible, three lines to wire up: &lt;a href="https://gateway.rrcsolutions.me" rel="noopener noreferrer"&gt;gateway.rrcsolutions.me&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I'm the founder and I read every piece of feedback. If you point it at a workload, I'd genuinely love to hear what your cache hit rate looks like — that's the number I obsess over.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>openai</category>
      <category>webdev</category>
    </item>
  </channel>
</rss>
