<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Aakash Chavan Ravindranath, Ph.D on Medium]]></title>
        <description><![CDATA[Stories by Aakash Chavan Ravindranath, Ph.D on Medium]]></description>
        <link>https://medium.com/@craakash?source=rss-cdc9df65fcc4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dVYecBF7vNSz7ZhU2tkxEw.png</url>
            <title>Stories by Aakash Chavan Ravindranath, Ph.D on Medium</title>
            <link>https://medium.com/@craakash?source=rss-cdc9df65fcc4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 03:59:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@craakash/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Parking is 38% of an EV charging stop in the Netherlands]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/parking-is-38-of-an-ev-charging-stop-in-the-netherlands-96ac12559f33?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/1512/0*Fo7AVfI3GBADZRav.jpg" width="1512"></a></p><p class="medium-feed-snippet">Two public registers, joined for the first time, show that the price on the charge point is only part of the bill. In Delft and Utrecht it&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/parking-is-38-of-an-ev-charging-stop-in-the-netherlands-96ac12559f33?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/parking-is-38-of-an-ev-charging-stop-in-the-netherlands-96ac12559f33?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/96ac12559f33</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[website]]></category>
            <category><![CDATA[netherlands]]></category>
            <category><![CDATA[parking]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 10:19:35 GMT</pubDate>
            <atom:updated>2026-10-07T10:20:01.651Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[python -m Is the Flag Most Python Devs Ignore (and It Fixes Your Import Errors)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">Two characters stand between ModuleNotFoundError and a working package.</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/python-m-is-the-flag-most-python-devs-ignore-and-it-fixes-your-import-errors-dbe6ca69793d?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/python-m-is-the-flag-most-python-devs-ignore-and-it-fixes-your-import-errors-dbe6ca69793d?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/dbe6ca69793d</guid>
            <category><![CDATA[design-systems]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[package]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 15:03:09 GMT</pubDate>
            <atom:updated>2026-09-28T15:03:09.958Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Built a Claude Code Model Router to Cut Token Costs. Then I Measured It.]]></title>
            <link>https://medium.com/@craakash/i-built-a-claude-code-model-router-to-cut-token-costs-then-i-measured-it-c20ee11cae3a?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/c20ee11cae3a</guid>
            <category><![CDATA[tokenmaxing]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[stories]]></category>
            <category><![CDATA[optimization]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Wed, 16 Sep 2026 08:06:13 GMT</pubDate>
            <atom:updated>2026-09-16T08:23:40.577Z</atom:updated>
            <content:encoded><![CDATA[<h3>A weekend plugin, eight controlled test runs, and a breakeven number that made me use it a lot less than I planned.</h3><p>My Claude Code usage kept running out mid-afternoon. I run the data team at an e-bike drivetrain company by day and a small analytics consultancy on the side, so I have two kinds of sessions: long agentic builds (a demand-forecast engine, a NetSuite tooling extension) and a pile of small stuff (monthly market-share emails, SKU cleanups, “which file does X”). Paying Fable 5.1 prices for the small stuff felt stupid.</p><p>So I did what everyone does: decided to build a router that sends easy work to a cheap model and keeps the expensive one for hard problems. I got a working plugin out of it. I also got two numbers I didn’t expect, and they changed how I use the thing. This post is mostly about the numbers.</p><h3>First, what Claude Code actually lets you control</h3><p>I wasted an evening on this, so here it is up front. I went through the <a href="https://code.claude.com/docs/en/model-config">model config</a>, <a href="https://code.claude.com/docs/en/sub-agents">subagents</a>, <a href="https://code.claude.com/docs/en/skills">skills</a> and <a href="https://code.claude.com/docs/en/hooks">hooks</a> docs looking for a way to switch the session’s model from code. There isn’t one.</p><p>/model is user-only. Hooks can inject context, rewrite a tool call&#39;s input, or block things, but not touch the model. There&#39;s even a PreModelSwitch hook, and all it can do is veto a switch you made yourself.</p><p>What you <em>can</em> set is the model each <strong>subagent</strong> runs on (frontmatter in agents/*.md, or the model parameter on the Agent tool) and the model a <strong>skill&#39;s</strong> own turn runs on (frontmatter in SKILL.md). Anthropic&#39;s own /model opusplan is built on exactly this idea: Opus while planning, Sonnet while executing.</p><p>So any “router” for Claude Code is really: keep the session on the cheap model, and steer which subagent gets which job. Once I understood that, the design fell out.</p><h3>The plugin</h3><p>I called it tiering. Three tiers, a classifier, four agents, two hooks. About 400 lines of stdlib Python and a handful of markdown files.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4fhqi2361ZpyFCWi1V1EUw.png" /></figure><p><em>How one turn flows. Blue is hooks, orange is models, green is the usage log.</em></p><p><strong>The classifier</strong> runs as a UserPromptSubmit hook on every prompt. It&#39;s deliberately dumb: keyword lists for hard work (root cause, architect, refactor across, a stack-trace regex, more than three file paths), keyword lists for trivial work (rename, format, which file), a never-downshift list (auth, secret, prod, delete, payment never go below the middle tier), and a rule that if you name a model in the prompt, that wins. If the score lands in a grey zone it can make one cheap Haiku API call to break the tie. Whatever it decides gets injected as context:</p><pre>TIERING: SUGGESTED_TIER=T2 ROUTE=fable-solver MODEL=fable<br>REASON=&quot;hard cue(s): root cause; stack trace&quot; (heuristic).<br>This is a recommendation, not a command.</pre><p>That last line matters. Sonnet is a better classifier than my regexes. The hook just makes the cheap default the <em>default</em>.</p><p><strong>The agents</strong> are four markdown files, each pinned to a model:</p><pre>---<br>name: fable-solver<br>description: Top tier. Expensive, use deliberately. Architecture, root cause, cross-file refactors.<br>model: fable<br>effort: high<br>---</pre><p>haiku-worker is read-only and capped at four turns. sonnet-worker does routine sub-tasks. fable-solver gets a self-contained brief for the hard stuff. fable-reviewer is read-only and returns a verdict.</p><p><strong>The usage log</strong> hangs off the Stop and SubagentStop hooks. Claude Code writes per-turn message.usage into the transcript JSONL, so the hook reads the transcript incrementally by byte offset and appends rows to ~/.claude/tiering/usage.jsonl. /tiering:report prices those rows and tells you what you spent by model and what the same tokens would have cost on Fable.</p><p>/tiering:force fable pins a tier, /tiering:off turns the whole thing off. Every script fails open: an exception prints {} and exits 0, so a bug in my hook can&#39;t block your turn. 133 unit tests, CI on two Python versions, because I&#39;ve been burned by hooks before.</p><p>Install, if you want to try it:</p><pre>/plugin marketplace add craakash/claude-tiering<br>/plugin install tiering@aakash-plugins<br>/model sonnet</pre><h3>Finding one: escalation isn’t free, and the breakeven is low</h3><p>Here’s the assumption every routing post makes without saying it: sending a hard task to a stronger model costs what that model charges for the task. It doesn’t.</p><p>When Fable answers you inline in a long session, most of its input is a <strong>cache hit</strong>. Fable 5.1 bills cache reads at $0.25 per million tokens versus $10 for fresh input. A typical agentic turn for me is around 2k fresh tokens in, 40k cache read, 1.5k out. That’s about $0.105.</p><p>A Fable <strong>subagent</strong> starts cold. There is no cache. It has to re-read whatever you brief it on and then write a full report back. Call it 15k in, 3k out: roughly $0.30. Three times the inline turn, and the Sonnet orchestrator turn that called it is on top.</p><p>I modelled a 100-turn session at <a href="https://platform.claude.com/docs/en/about-claude/pricing">list prices</a> with 20% trivial turns and varied the share of hard turns:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2lcEv2-aIXCyvoAwICKpyw.png" /></figure><p><em>Above about 26% of turns escalated, the “cheap session plus escalation” line crosses the all-Fable line.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mm97fZIJj89Q0elZRMcqpg.png" /></figure><p>Two things I took from this.</p><p>The router’s real job is to make Sonnet <em>safe to stay on</em>. Plain /model sonnet is the 74% saving. The router is the escape hatch that stops you flipping back to Fable out of nerves.</p><p>And for the genuinely hard sessions, turn it off. My forecasting build has an architecture question every third turn. That’s a Fable session. Escalating it thirty times through cold subagents is the most expensive way to run it.</p><p>The exact breakeven will move with your cache hit rate and how much the subagent has to re-read. The shape won’t. Cold context is expensive, and a router that pretends otherwise overspends precisely on the turns you built it for.</p><h3>Finding two: terse output barely moves the token count</h3><p>There’s a genre of “make Claude talk like a caveman” skills claiming 60 to 75% token savings. I had one installed. Here’s where the tokens in a single Fable turn actually go:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c-g0EDPydtRCSzCGdFny9A.png" /></figure><p>Output is about 3% of tokens. Prose to me is a slice of that; the rest is code, tool calls and thinking. Cut the prose in half and you’ve saved maybe 1% of tokens. In dollars it’s better, because output is priced at 5x input, but it’s still single digits. The 75% figures are 75% of the prose. The prose isn’t the bill.</p><p>If what you care about is a subscription meter that runs out slower, verbosity is the last lever, not the first.</p><h3>Finding three: what did save tokens, measured</h3><p>I wrote a short set of rules I’ve been calling lean mode. Most of them are about <em>reading</em>, not writing:</p><ul><li>Read line ranges, not files. offset/limit on Read, head_limit on Grep.</li><li>Send search and multi-file exploration to an Explore subagent. Ask for a conclusion, not file contents.</li><li>Don’t re-read a file you just edited. Don’t paste tool output back into prose.</li><li>Batch independent tool calls. One Bash call that does five things beats five calls.</li><li>Whole task in one pass. One clarifying question at most.</li><li>Answer first, five lines of prose unless asked for more, only the changed code.</li><li>Set model explicitly on every delegation.</li></ul><p>Then I tested it properly, or as properly as an evening allows. Eight realistic tasks against the plugin’s own repo. Each task run twice on <strong>Sonnet</strong> with an identical prompt: once plain, once with the rules prepended. Same repo state, same model, nothing else different.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Hm9qfu300tnpFdIZnzLuzA.png" /></figure><p><em>600,810 tokens down to 536,041. Just under 11%.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qnsnfMe2gVa14h9WjoBmhw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n4O9P-LPxjrrbBWAqesl6w.png" /></figure><p>What the table says, as I read it:</p><p>The saving lives in exploration. “Find every hardcoded string” dropped 22% because the lean run handed the search to one Explore agent and got a conclusion back, instead of running eleven greps and holding all the output in context.</p><p>Pure generation doesn’t move. Writing a test file: 1%. The tokens are the work.</p><p>The multi-file rename saved 19% with <em>more</em> tool calls (65 versus 61), because the lean run stopped re-reading files after editing them. Bytes per call matter more than call count.</p><p>And there’s a quality cost I want to be honest about. The lean run wrote 9 tests where the default wrote 17. On the --json task, lean changed a function signature and had to patch three tests; the default kept it backward compatible. The rules say correctness beats brevity. The model mostly listens. Mostly.</p><p>Caveats: n=8 is a direction, not a statistic; any single task swings ±10% between runs. The subagents ran without a real session’s baggage (system prompt, MCP tool schemas, memory files), so in a live session I’d expect the context rules to matter more, not less. And the breakeven above is a model at list prices. /tiering:report exists so you can replace it with your own number, which is what I&#39;m doing now.</p><h3>What I’d actually tell a colleague</h3><p>In the order I’d do them:</p><ol><li>Put &quot;model&quot;: &quot;sonnet&quot; in ~/.claude/settings.json. One line. Roughly five times cheaper per token than Fable.</li><li>Keep context small. /clear between unrelated tasks, /compact before it&#39;s forced, read ranges not files, delegate exploration. This is the 11% and it compounds every turn.</li><li>Disconnect MCP servers you aren’t using today. Every tool schema rides along on every turn whether you call it or not.</li><li>Drop effort to medium for routine work. High-effort thinking on a rename is money on fire.</li><li><em>Then</em> a router, if you still want one, as an escape hatch that lets you stay on Sonnet. Mind the breakeven. Or just use /model opusplan, which is the two-tier version Anthropic ships.</li><li>Verbosity last. Real, small, and the one everybody does first.</li></ol><p>I’ve believed for a while that the harness around a model matters more than the model. This was a small, slightly humbling confirmation: the harness is mostly what you let into context, and a router that ignores cache economics can cost more than the problem it solves. Measure yours before you trust it.</p><h3>Code</h3><p>Plugin, tests, CI and the lean-mode rules: <a href="https://github.com/AnalyticsAscentAkx/claude-tiering">github.com/craakash/claude-tiering.</a></p><p>Prior art I found <em>after</em> building mine, which is its own lesson: <a href="https://agentskill.sh/plugins/0xrdan/claude-router">claude-router</a> does similar routing with an Opus ceiling and a stats dashboard. Don’t run two UserPromptSubmit classifiers at the same time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c20ee11cae3a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Use Claude in the Terminal and I’m Never Going Back. But One Thing Was Missing.]]></title>
            <link>https://medium.com/@craakash/i-use-claude-in-the-terminal-and-im-never-going-back-but-one-thing-was-missing-f5fac6c0ced5?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/f5fac6c0ced5</guid>
            <category><![CDATA[github]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[terminal]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[pipy]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Mon, 11 May 2026 15:33:10 GMT</pubDate>
            <atom:updated>2026-05-11T15:36:47.847Z</atom:updated>
            <content:encoded><![CDATA[<h4>So I found undocumented Anthropic API headers and built a tool to fix them.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/558/1*01jdsh1Dv5nimnrrrY2nCg.png" /></figure><p>I’ve been using Claude Code in the terminal for a while now, and honestly, I can’t imagine going back to the desktop app for actual work.</p><p>Don’t get me wrong — the desktop app is fine. But once you start using Claude from the terminal, you realise how much faster everything feels. Not because the model is faster or uses fewer tokens (it’s the same under the hood), but because of how you work with it.</p><h3>Why I Prefer the Terminal</h3><p><strong>You can pipe anything into it.</strong> That’s the big one. Want Claude to review your recent changes?</p><pre>git diff main | claude -p &quot;review this for bugs&quot;</pre><p>Want it to analyze logs?</p><pre>tail -500 server.log | claude -p &quot;find anomalies&quot;</pre><p>Try doing that in a desktop app. You’d be copy-pasting for 10 minutes.</p><p><strong>It fits into your existing workflow.</strong> I’m already in the terminal, running git commands, spinning up servers, checking deployments. Claude Code is just another command. No switching windows, no losing context, no clicking through menus.</p><p><strong>It works everywhere.</strong> SSH into a remote server? Claude works. CI/CD pipeline? Claude works. Headless VM? Claude works. Desktop app? Good luck.</p><p><strong>Scripting and automation.</strong> You can chain Claude into shell scripts, use flags like ` — continue` to pick up where you left off, ` — model` to switch models on the fly. The terminal version is composable in ways the desktop just isn’t.</p><p><strong>And honestly?</strong> I just think faster in a terminal. There’s less visual noise. No sidebar, no settings panel, no notification popups. Just me, my code, and Claude.</p><h4>The One Thing That Was Missing</h4><p>So here I am, happily working in the terminal, shipping features, debugging at 2 AM — and then that thought hits:</p><p><strong>”Am I about to hit my usage limit?”</strong></p><p>And what do I do? I stop everything. Open the Claude desktop app. Click settings. Navigate to usage. Stare at a progress bar that says “50% used, resets in 22 hours.” Go back to my terminal.</p><p>The irony was not lost on me. I moved to the terminal to avoid context-switching, and the one thing that keeps pulling me back to the desktop is a simple progress bar.</p><p>I did this probably 10 times a day. And every single time it annoyed me a little more.</p><h4>Looking for Solutions</h4><p>First thing I did was look for existing tools. I found `claude-monitor` — it parses your local session files, counts tokens, runs ML predictions, calculates P90 percentiles, and gives you an *estimate* of your usage.</p><p>Keyword: <strong>estimate</strong>.</p><p>It’s guessing. It reads token counts from local files and tries to reverse-engineer what percentage of your plan that represents. Sometimes it’s close. Sometimes it’s way off. And it needs config files, plan selection, Pydantic models, Sentry integration — the whole thing felt like overkill for what I needed.</p><p>I just wanted the actual number. The same number I see when I click on `claude.ai/settings/usage`. That exact percentage. That exact reset time.</p><h4>The Discovery That Changed Everything</h4><p>I spent a while trying different approaches. Scraping the web dashboard? Cloudflare blocks it. Hidden API endpoint? Tried dozens — all 404. Local files? They have token counts but not plan percentages.</p><p>Then I had a thought — what if the data is already there, just hiding in plain sight?</p><p>I made a simple API call and dumped <strong>every single response header</strong>. And there it was:</p><pre>anthropic-ratelimit-unified-5h-utilization: 0.11<br>anthropic-ratelimit-unified-5h-reset: 1778517600<br>anthropic-ratelimit-unified-7d-utilization: 0.03<br>anthropic-ratelimit-unified-7d-reset: 1778972400<br>anthropic-ratelimit-unified-status: allowed</pre><p>I stared at my screen for a solid 30 seconds.</p><p>Anthropic sends your <strong>exact plan utilization</strong> as response headers on every single API call. The 5-hour window (that’s your “Current session” on the dashboard) and the 7-day window (that’s your “Weekly limits”). The exact percentages. The exact reset timestamps.</p><p>These headers are completely undocumented. Not in the API docs. Nobody talking about them online. But they’re there, on every response, just waiting to be read.</p><p>— -</p><h4>Building It</h4><p>Once I knew where the data was, the tool took maybe an hour. The approach is dead simple:</p><p>1. Read the OAuth token from macOS Keychain (Claude Code stores it there when you log in)<br>2. Make one tiny API call — 1 token to Haiku, costs literally nothing<br>3. Read the rate limit headers from the response<br>4. Draw progress bars in the terminal</p><p>That’s it. One file. About 200 lines. Zero dependencies beyond Python’s standard library.</p><pre>Plan usage limits Team<br>Current session<br> Resets in 4 hr 38 min<br> ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 13% used<br>Weekly limits<br>All models<br> Resets in 130 hr 58 min<br> ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 4% used<br>Last updated: 14:01:55 ● within limits</pre><p>Same data as the desktop app. Because it IS the same data, from the same source.</p><p>The /usage Slash Command — This Is the Real Win</p><p>The terminal command is nice. But here’s what actually made my day.</p><p>Claude Code has this feature called Skills — you create a markdown file, and it becomes a slash command you can use inside your session. So I created `/usage`.</p><p>Now when I’m in the middle of a conversation with Claude, debugging something complex, and I want to check my limits — I just type `/usage`. Right there. Same terminal. Same session. No context switch.</p><p>That was the whole point. I moved to the terminal to stop switching between apps. And now I finally don’t have to.</p><p><strong>The Undocumented Headers (For the Curious)</strong></p><p>For anyone building tools around Claude or just curious about what Anthropic sends back, here’s the full list:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UflTKGkghKy1WGzMlMHq3g.png" /></figure><p>These come back on every `/v1/messages` response. They work with OAuth tokens from Claude Code. I haven’t tested with regular API keys but I imagine they return something similar.</p><p>Fair warning — these are undocumented, so Anthropic could change them tommorow. But they’ve been consistent across every call I’ve made, and the numbers match the dashboard exactly.</p><h4>How It Compares to Existing Tools</h4><p>I want to be fair. `claude-monitor` is a solid project with way more features — burn rate analysis, cost tracking, ML predictions, Rich UI. If you want deep analytics on your token patterns, go use that.</p><p>But if you just want the answer to: ”<strong>How much of my plan have I used and when does it reset?</strong>”— that’s what this does.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fRHe8u7AAHGENFyOt80fWw.png" /></figure><h4>Try It</h4><pre>pip install claude-plan-usage<br>claude-plan-usage</pre><p>If you have Claude Code installed and logged in on macOS, it just works. Zero config.</p><p>GitHub: [github.com/craakash/claude-plan-usage](<a href="https://github.com/craakash/claude-plan-usage">https://github.com/craakash/claude-plan-usage</a>)</p><p>Currently, macOS only supports this because the auth token lives in Keychain. Linux and Windows PRs are very welcome.</p><h4>What I Learned</h4><p>Sometimes the data you need is already there — you just have to look in the right place. I spent hours trying to reverse-engineer local files and find hidden API endpoints. The answer was sitting in the HTTP headers of every API call I was already making.</p><p>And if something annoys you enough times, just build the thing. It took an evening. Now I never have to open the desktop app again.</p><p><em>If this was useful, give the [repo](</em><a href="https://github.com/craakash/claude-plan-usage"><em>https://github.com/craakash/claude-plan-usage</em></a><em>) a star. And if you discover other undocumented headers hiding in there, let me know — I’m curious what else Anthropic isn’t telling us.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f5fac6c0ced5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[crypto-yfinance: A Python Library for Cryptocurrency Data (Like yfinance, but for Crypto)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/crypto-yfinance-a-python-library-for-cryptocurrency-data-like-yfinance-but-for-crypto-3f7056d3f11e?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/600/1*n8rzUeBytaiERMBFHVaM9g.png" width="600"></a></p><p class="medium-feed-snippet">If you&#x2019;ve ever worked with stock data in Python, you probably know and love yfinance. It makes financial data simple. But in the crypto&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/crypto-yfinance-a-python-library-for-cryptocurrency-data-like-yfinance-but-for-crypto-3f7056d3f11e?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/crypto-yfinance-a-python-library-for-cryptocurrency-data-like-yfinance-but-for-crypto-3f7056d3f11e?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f7056d3f11e</guid>
            <category><![CDATA[python-libraries]]></category>
            <category><![CDATA[personal-growth]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[personal-finance]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Thu, 21 Aug 2025 17:40:10 GMT</pubDate>
            <atom:updated>2025-09-03T13:32:18.188Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[11 Best Crypto Data APIs and Python Libraries for Reliable Market Data in 2025 (Free & Paid)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/11-best-crypto-data-apis-and-python-libraries-for-reliable-market-data-in-2025-free-paid-50d369f52876?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/2400/1*krjHi1j7KXhBbya3-mTTlw.png" width="2400"></a></p><p class="medium-feed-snippet">Discover the 11 best crypto data APIs and Python libraries in 2025. Free vs. paid tiers, real-time data, on-chain metrics, and code&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/11-best-crypto-data-apis-and-python-libraries-for-reliable-market-data-in-2025-free-paid-50d369f52876?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/11-best-crypto-data-apis-and-python-libraries-for-reliable-market-data-in-2025-free-paid-50d369f52876?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/50d369f52876</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Wed, 20 Aug 2025 10:50:03 GMT</pubDate>
            <atom:updated>2025-08-20T10:50:03.941Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[9 Best yfinance Alternatives for Reliable Financial Data in 2025]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/9-best-yfinance-alternatives-for-reliable-financial-data-in-2025-b48b4aaa9ac9?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/2379/1*OyHLjUJYEjcBZTHIUHx_DQ.png" width="2379"></a></p><p class="medium-feed-snippet">yfinance has been the go-to library for Python developers pulling market data. It&#x2019;s quick, free, and easy to use&#x200A;&#x2014;&#x200A;but also fragile.</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/9-best-yfinance-alternatives-for-reliable-financial-data-in-2025-b48b4aaa9ac9?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/9-best-yfinance-alternatives-for-reliable-financial-data-in-2025-b48b4aaa9ac9?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/b48b4aaa9ac9</guid>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[api]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Thu, 14 Aug 2025 19:05:03 GMT</pubDate>
            <atom:updated>2025-08-14T19:05:03.349Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Hidden Cost of Premature Rounding: A Quantitative Analysis from Arithmetic to Machine Learning]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/the-hidden-cost-of-premature-rounding-a-quantitative-analysis-from-arithmetic-to-machine-learning-fcdcb074062a?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/901/1*yzXzLixl3z_CkiE3lUssbw.png" width="901"></a></p><p class="medium-feed-snippet">Premature rounding silently corrupts results in finance, statistics, and machine learning. This article quantifies the error growth, shows&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/the-hidden-cost-of-premature-rounding-a-quantitative-analysis-from-arithmetic-to-machine-learning-fcdcb074062a?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/the-hidden-cost-of-premature-rounding-a-quantitative-analysis-from-arithmetic-to-machine-learning-fcdcb074062a?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcdcb074062a</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[floating-point-precision]]></category>
            <category><![CDATA[numerical-analysis]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[rounding-errors]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Tue, 05 Aug 2025 20:50:28 GMT</pubDate>
            <atom:updated>2025-08-05T20:50:28.536Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[How to Read and Extract Data from PDF Files: A Full Working Example Using Python and OpenAI]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/how-to-read-and-extract-data-from-pdf-files-a-full-working-example-using-python-and-openai-3629da81834f?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/1024/1*PCiH5CrkI4DNY7tdg3oqVA.png" width="1024"></a></p><p class="medium-feed-snippet">Meta Description: Learn how to read and extract data from PDF files using Python, PyMuPDF, and OpenAI&#x2019;s GPT. This guide presents a&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/how-to-read-and-extract-data-from-pdf-files-a-full-working-example-using-python-and-openai-3629da81834f?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/how-to-read-and-extract-data-from-pdf-files-a-full-working-example-using-python-and-openai-3629da81834f?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3629da81834f</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[pdf]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Wed, 30 Jul 2025 13:25:03 GMT</pubDate>
            <atom:updated>2025-07-30T13:25:03.506Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[How the 20 Questions Toy Outsmarted Millions — and How You Can Build Your Own AI Game]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@craakash/how-the-20-questions-toy-outsmarted-millions-and-how-you-can-build-your-own-ai-game-ebe215fbe154?source=rss-cdc9df65fcc4------2"><img src="https://cdn-images-1.medium.com/max/600/1*5ddyOjhz4zfeENWqN7Sunw.png" width="600"></a></p><p class="medium-feed-snippet">If you grew up in the late 1990s or early 2000s, chances are you encountered a quirky little plastic gadget that felt almost magical in&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@craakash/how-the-20-questions-toy-outsmarted-millions-and-how-you-can-build-your-own-ai-game-ebe215fbe154?source=rss-cdc9df65fcc4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@craakash/how-the-20-questions-toy-outsmarted-millions-and-how-you-can-build-your-own-ai-game-ebe215fbe154?source=rss-cdc9df65fcc4------2</link>
            <guid isPermaLink="false">https://medium.com/p/ebe215fbe154</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[information-theory]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Aakash Chavan Ravindranath, Ph.D]]></dc:creator>
            <pubDate>Wed, 16 Jul 2025 11:56:57 GMT</pubDate>
            <atom:updated>2025-07-16T11:56:57.402Z</atom:updated>
        </item>
    </channel>
</rss>