DEV Community: sandeep mudhi raj

A family business was spending over ₹15,000 a year on its website. Today, it spends just ₹2,500.

sandeep mudhi raj — Tue, 27 Oct 2026 15:44:01 +0000

Whenever I write about AI or automation, my focus is cost-cutting: more output, fewer tokens, lower bills. I follow this in my own work too. Xel Studio and SearchWala both run at zero cost.

Recently I applied the same approach to a family business in the finance space. Its website serves investors and traders. The link is in the first comment if you want to see it.

Here is what changed.

Hosting cost up to ₹8,000 a year, mostly for services nobody used. I moved the site to Netlify. Cost now: zero.

Every small change meant paying a developer again. That dependency is gone.

More automation and workflows are coming to this site. Normally that means expensive business plans. Instead, I split the work across free platforms. Netlify is the face, the frontend. GitHub and GitLab are the brain, running the backend tasks through their free automation runners. Each free plan has its own monthly limit, so I split the work between them, and neither one runs out.

I also rebuilt the frontend. A design like this would cost over ₹15,000 in the market, and I haven't even counted that.

The old plan expires in a few days, and then the domain comes fully under our control. The only bill left is the ₹2,500 domain fee.

Now, be honest. A live business backend running on GitHub and GitLab, platforms most people only use to store code. Smart engineering, or a clever workaround that will break one day?
Website link https://globaltekfinancialservices.netlify.app/

WebDevelopment #Automation #CostOptimization #Netlify

The Biggest Problem With Gemini 4 Argon. It Thinks Too Much.

sandeep mudhi raj — Wed, 21 Oct 2026 09:50:31 +0000

For the past few days I have been using Gemini 4 Argon and watching it closely. The model itself is good. It gives solid answers and gets the work done. But it has one problem that hurts both the user and the company.

The problem is overthinking. Whether you run it on low, normal or high, and whether the question is small or big, it thinks far more than it needs to. I simply asked it to visit a website and tell me what was on it, or to check an email GitHub had sent to my Gmail. A Flash model does the same job quickly and correctly. But Argon thinks so much that tokens get wasted for no reason. It does not understand the intent behind the question and answer accordingly. It just keeps thinking on its own.

Take a simple example. You ask a straightforward question and the model thinks for a full minute. That causes two losses. First, Google's servers carry load they did not need to. Second, say your usage limit was at 10 percent. That one question pushes it straight to 20 or 25 percent. The tokens are gone, and the answer is the same one you could have got much faster.

Almost every model used to have this problem. OpenAI's GPT models have reduced it a lot, and in my view Claude handles it best, whether it is Opus or Sonnet. I use Claude Opus 5.5 regularly, and for work like software it can compete with Gemini. Even on max effort, if the question is small, it answers right away. It can think deeply too, but it knows when to think and how much, just like we humans do.

One thing I should be clear about. So far I have only used Argon on mobile, not for development, because my current plan does not include it and paying extra for it right now did not seem worth it. If I need it later, I will test it more deeply.

A good model is not the one that thinks the most. A good model is one that gets our work done well at a lower cost. When a model thinks only as much as it needs to, the user saves their usage limit, the company saves server resources and money, and the answer arrives faster. Both sides win.

Building an autonomous AI agent is easy. Keeping it alive 24/7 for four months on a ₹0 budget is the real engineering problem

sandeep mudhi raj — Sun, 18 Oct 2026 09:39:50 +0000

.My first version taught me that the hard way. One LLM, no fallback. Every rate limit or outage froze the entire pipeline. Image generation was worse: I went through several free tiers before Cloudflare, and Hugging Face's free inference simply wasn't dependable enough for a system that never sleeps.

So I rebuilt it as XeL Studio, an open-source, self-driving newsroom. Every 30 minutes it finds tech news, writes the article, generates a cover image, and publishes. Once a day it cleans its own database. No human in the loop.

Here's how I designed around free-tier limits instead of fighting them:

Two LLMs, one pipeline
Groq and Cerebras run behind a fallback router. If one hits a rate limit or goes down, the other takes over. That single change fixed most of v1's failures.

Vercel for the frontend, GitHub Actions for the heavy lifting
The Next.js + TypeScript frontend is hosted on Vercel. But serverless timeouts kill long jobs, so AI generation and Python scraping run on GitHub Actions as a scheduled backend.

Three databases, each with one job

Firebase: frontend reads. I built the UI layer around it because the load speed is excellent.
Supabase (PostgreSQL): long-term article storage on its own free tier, so the archive can keep growing. It also runs the daily cleanup.
JSON: the news pipeline only. When an AI is reading and writing structured data, JSON has been the most reliable format for me.

Discovery via Searchwala
News discovery runs on Searchwala, another open-source project of mine.

Media at ₹0
Cloudflare generates the cover images. Cloudinary stores and delivers them.

Honestly, this is my favourite part of building: taking free tiers that are weak on their own and combining them into something that holds up in production.

Both XeL Studio and Searchwala are open-source. Links in the first comment.

So, would you trust a free-tier stack like this in production, or is it a ticking time bomb?

OpenSource #AIAgents #SystemDesign #Serverless #BuildInPublic

Website link:
https://xel-studio.vercel.app/
GitHub link:
https://github.com/SandeepAi369/xel-studio

Recently Released: 'Agent First Browse' The Autonomous AI Web Agent

sandeep mudhi raj — Mon, 12 Oct 2026 10:14:04 +0000

A massive shift has arrived in the world of web automation. You no longer need to open a browser to manually click, fill out forms, or search for products. I recently released Agent First Browse on GitHub—an autonomous AI web agent that transforms your simple text prompts into real-time browser actions.

Whether it's finding the best deal on Amazon and adding it to your cart, navigating to a website to fill out a complete form, or autonomously publishing posts on platforms like X (Twitter) and Reddit—you just need to give it a command.
For example, you can simply prompt: "Open the browser, go to HackerNews, find the top three AI articles, and give me a quick summary." The agent will open the browser, complete the entire task for you, and hand over the final report.

How It Works & Key Features

Instead of simple scripts, this agent uses a highly smart backend giving it complete freedom to think and act like a human. It operates with its own virtual monitor, automatically moving the mouse and typing on the keyboard completely on its own. Just give a command once, and it does all the work.
It acts like a smart problem solver. If an unexpected error occurs—like a promotional pop-up blocking a submit button—it instantly recognizes the hurdle, autonomously closes the pop-up, and proceeds with the task.
Because it is given full freedom to think deeply and make the right decisions, it consumes slightly more tokens but delivers highly accurate results.
The complete deep-dive into how its AI brain and nervous system work (using LLMs, LangGraph, MCP, Python, and Playwright) across its massive 98 files and 25,000-line codebase is thoroughly documented on the GitHub page.

Important Note (Security & Testing Phase)

This is currently in the Testing Phase and runs perfectly for Linux and CLI users.
As I haven't focused deeply on security yet, please strictly avoid using it for sensitive tasks (like banking, payments, or passwords). Use it for testing and report bugs.

Future Updates

Just as I update my other projects like SearchWala and XeL Studio from time to time, I will consistently update this agent as well, including full compatibility for Windows users.
For working proof, a real video of Amazon automation is attached in the repository.

Source Code & Setup Guide

Everything you need to get started is well-documented. Detailed, step-by-step instructions on how to clone the files, install dependencies, configure API keys, and run the agent are provided on the GitHub page.
You can check out the complete source code, contribute, or start testing it yourself here:

GitHub: https://github.com/SandeepAi369/Agent-first-browse