Building a Production-Grade Web Platform Using Free-Tier AI Tools
Introduction: Why This Project Mattered
If you've hit this page, you're probably interested in how AI can transform how we create applications in this day and age. Guess what? We built 100% of this site you are currently viewing using AI while possessing only the very basic HTML web development knowledge. This site, localized for the Kenyan market, utilizes Python on the backend and HTML on the front end. This was an ambitious project to create a unique application for the Kenyan market while testing the boundaries of vibe coding.
Brief Overview of the Project
This is not a basic website. It's a fully featured business utility platform. Registered users can create WhatsApp/Instagram style stories with the ubiquitous reactions and replies functionality. In the main search page, search is powered by AI making searching for products easy and fun. You can search for products using natural language e.g., "Best TV for my grandma". The AI agent in the background understands what you are searching for, analyzes your request, checks what televisions are the best fit for your use case, and recommends a number of televisions while giving you its reasons behind its choices.
Users who decide to register as vendors have a whole menu of additional features they can utilize including posting classified products (products have inbuilt Rich Snippets for better SEO), RFQs, create quotations, invoices, delivery notes and receipts, buy and send bulk SMS to contacts they create, who can be segmented into groups. There is an activity feed page where public activities are lined up. Other features include wishlists, various analytics, messaging, support tickets and blogs.
🎮 Gamification Feature
The platform includes a gamification feature where users/vendors are awarded points for actions on the site including likes, shares, reposts, putting up RFQs etc. Further fine-tuning of the gamification system will be done to include what rewards a user can get from the points accumulated.
The Challenge: Building a Production-Grade Platform with Zero LLM Spending
Building a web application like this one using AI requires some homework. You have to know some of the strengths and weaknesses of popular AI LLMs. In our case we used Anthropic's Claude AI as the main coding agent, and ChatGPT, Qwen and Google's Gemini for supporting roles. One thing to note is we used the free tier for all the LLMs.
Target Users and Use Cases
A classifieds portal is something we were always interested to have. But the competition in Kenya is enormous. There are a few top classifieds sites in Kenya commanding a big market percentage, and numerous other smaller sites taking up the rest of market share. So the big question is, how can we convince a vendor to sign up to yet another classifieds site?
The answer was providing what other classifieds sites don't. A one-stop platform where vendors can utilize various useful features to enhance their user experience and business needs. That's where features like quotations, invoices, RFQs, ticketing and some social media-like features come into play. Other exciting features, for example unified communications, KRA tax compliance, jobs, funzone, and many more are planned in future releases.
Why Free-Tier LLMs?
💰 Cost Realities for Startups and Small Teams
With zero development budget, paying monthly LLM subscription was not an option. We had to use what we have: internet and the will to build something truly usable to launch our online business.
Comparison: Paid APIs vs Free-Tier Experimentation
During the development process, we got a surprise offer of $250 worth of trial for Claude Pro through which we had access to more Claude models, usage of Claude Code, and allegedly more usage than the free tier. More usage would mean higher token limits etc.
We utilized this offer for other experiments. Our conclusion is that where the Pro tier truly shines is with Claude Code. But with the normal Claude AI, there are weekly token limits which are not there, or at least we never hit them, in free tier. In free tier, when you hit token limits, you get a timeout of something like 4 hours before you can continue with your project.
With this cycle, you can progress with your project step by step. But with the paid Claude Pro, we hit weekly token limits within the same day and had to wait for a whole week before continuing with the project. This was quite frustrating. Our conclusion, at least as far as our projects are concerned, and with our limitation of the deep workings of Claude Pro, is that free tier was better than paid tier as far as Claude is concerned.
LLM Selection Strategy
Criteria for Choosing LLMs
We did not do any kind of thorough research on which LLM to use before starting the project. Before starting the shopping portal project, we had done some smaller scale experimentation projects, building some modules/plugins for some well-known open source platforms, which basically ended at experimentation stage. But we learned some important points that helped with this project:
🔑 Key Insights
- Claude excels at long documents, while ChatGPT and Gemini balance context with speed
- Rate limits matter for scale: free tiers of ChatGPT, Gemini, Qwen, and Claude vary widely
- Qwen was quite good at debugging well-logged errors while poor at debugging silent errors (errors that do not appear on logs)
- You cannot, and I repeat, you cannot use free tiers of Google Gemini, ChatGPT and Qwen to build any big project like our shopping portal. But you can with Claude and the results speak for themselves
How We Mixed Multiple Free-Tier Models
As we said earlier, Claude was our main coding agent while Qwen, ChatGPT and Gemini played supporting roles. We mostly used Qwen, ChatGPT or Gemini for debugging specific errors. Of course the first choice for debugging is the person with the bird's eye view of the whole project, which is Claude. But sometimes, either you have run out of tokens or Claude seemed unable to debug the error. In that case we pasted the error to the other LLMs to find a solution.
Prompts & Project Design
Creating the platform started with creating a projects folder in Claude, where you write the name and description of the project. You can also add special instructions for reuse across the whole project across different chats. Before starting the project, we discussed with Claude about our needs and it gave us a comprehensive overview and project insights, structure, tech stack and implementation roadmap. The roadmap is in phases. Commencing the project was simple. Once we were satisfied with the roadmap, it was a simple prompt: "start phase 1".
Phase one would be an MVP (Minimum Viable Product). This is a barebones codebase that can be installed and works but without any bells and whistles. One thing to note is that we did not develop this project on localhost. Right from the first page, the index page, was live on the net. Of course this is not the best way to develop an application. But it was exciting to see our project progress live on the web. Bugs would bring the whole site down but we were not worried. After all there was no traffic to the site as we developed it.
⚙️ Development Workflow
From there it was routine work. Prompt, prompt, prompt. Files are created, we upload them, test, debug, rinse and repeat. More and more features materialize. With any single chat, there is a maximum length after which you have to start a new chat.
To prevent LLM "forgetfulness" when starting a new chat, working within a project helps. But that is not all. Start by summarizing what you were working on before you reached maximum chat limits. This will help with seamless continuation of the feature you were working on. It's not foolproof though. Mistakes still happen. Sometimes updating a file can lead to bugs. If you simply copy paste the produced files, you may sometimes find some code blocks have been removed by the LLM setting you back. Sometimes it's one step forward, one step back. This behavior is least in Claude and most in Qwen. Never copy paste a Qwen-produced file without thorough checking.
Impact & Results
After a few weeks of work, the results are what you see here. A unique, fully featured shopping portal/business utility platform. It's open for business and works as expected. Of course it still has bugs to be ironed out, which we are sorting out as we find them. The codebase has not been reviewed by an experienced developer. We are sure that it may have kinks here and there. But the overall lesson is you can create a web application with zero coding knowledge using AI.
🚀 The Bottom Line
If we can build a fully featured shopping and business utility platform at zero LLM cost, you can prototype your idea today. Start small, test aggressively and don't let constraints pull you down.
Sign up and put xtra.co.ke to the test. Explore our portal, try the features, and see how far free-tier LLMs can go in a real product. Your feedback will help us refine the experience, improve reliability, and decide what to build next. Whether you're a founder, developer, or curious user, your insights matter. Join us, test the platform, and help shape what this becomes. It's production grade so you can also post your products and use the other features that are available.