4 min read
How I Created JiakAI Bot with Less Than $10 - Vibecoding
Built JiakAI — an AI-powered calorie-tracking Telegram bot — in half a day using vibecoding, all for under $10.
- AI
After hearing so much about people vibecoding and all the new AI power tools from the big tech companies, I decided to jump on the trend during my work break — while chilling at a coworking space with some cats. My goal was to try out Cursor with Claude Code, and build something fun, functional and fast with AI assistance. The result? JiakAI - a Telegram bot that lets you snap a photo of your meal and instantly get calorie and macro estimates.
It took less than half a day to build at a cost under $10 USD (you could probably do it cheaper if you know what you're doing), and it was entirely coded by AI while I was chilling with the cats. This is a light project, but the process taught me a lot about the power of AI-driven prototyping.
What Does JiakAI Do?#
JiakAI is a personal food-tracking assistant Telegram bot that supports:
- 📝 Text Input: Log what you eat by simply typing it in.
- 📸 Photo Input: Snap and send a meal photo directly via Telegram.
- 🤯 AI Vision: Uses OpenAI's GPT-4 Vision to identify food items from the photo.
- 🔥 Nutrition Analysis: Retrieves estimated calories, protein, carbs, and fat using the Nutritionix API.
- 📊 Daily Summary: Tracks your meals throughout the day and provides a summary on demand.
- 📂 Storage: Saves everything in Firebase so you can review your food logs anytime.
👉 For a quick demo of the bot, check out this LinkedIn post
How it’s built#
This project uses a simple but effective stack:
- Telegram Bot API (Python): Handles user interaction
- OpenAI Vision (GPT-4-Vision): Extracts food names from images
- Nutritionix API: Looks up nutritional values
- Firebase: Stores food logs per user
Cost Breakdown#
- OpenAI GPT-4-Vision API: ~$0.002 per image
- Nutritionix: Free for non-commercial use (within daily limits)
- Firebase (Firestore): Free tier used
How I Vibecoded#
I used Cursor with Claude Code installed. Here's my process:
- Drafting Ideas: I first drafted my app ideas and asked Cursor to create a Product Requirements Document (PRD) for me.
- Code Generation: Using the PRD.md, I then asked Claude Code (via Cursor terminal - you need to input your own Anthropic API key) to read the PRD and code the app based on the requirements.
- AI-Assisted Development: What was amazing is that you can even ask Claude Code in the terminal to help run the app for you. If you run into any setup or startup issues, Claude Code can even fix them for you - which means I touched zero code from the editor other than inserting the API keys in the .env file.
- Debugging: After testing, I ran into some runtime errors. Likewise, you can copy and paste the error to Claude to debug and fix it for you. I further added features such as access control to only allow selected Telegram IDs to access the functions (this is to reduce my OpenAI usage).
All I had to do was create accounts in Nutritionix, Firebase, and OpenAI, then paste the API keys. And I chose OpenAI Vision for image analysis simply because I had leftover credits from a previous project, no deep reasoning behind it.
The Source Code#
When I first built JiakAI, I wanted to share it openly so others could explore how AI-assisted development could create something functional in just a few hours.
Update (Oct 2025): The JiakAI repository was originally public for learning and experimentation. I’ve since made it private as I continue developing JiakAI into a monetized product.
JiakAI was originally prototyped entirely with AI assistance — mainly using Claude Code through Cursor. The entire app, from idea to working prototype, was generated by AI based on a PRD I drafted.
Since then, I’ve been refining and extending it into something more robust and production-ready — focusing on scalability, better architecture, and smarter calorie estimation models.
The bot still does what it set out to:
- 📝 Log meals through text
- 📸 Identify food from photos
- 🔥 Estimate calories and macros
- 📊 Track your daily nutrition summary
My Verdict on Vibecoding#
✅ Pros:
- Powerful for rapid prototyping - Perfect for quick POCs and testing out new technologies
- Great for experimentation - Helps you explore ideas without heavy upfront investment
❌ Cons:
- Not a replacement for real learning - You risk skipping over the deep understanding of architecture, security, and best practices
- Limited for production - Without proper knowledge, you might miss potential vulnerabilities or optimization opportunities
Key Takeaways#
Vibecoding is an awesome way to explore ideas and iterate fast. It lowers the barrier to creation, especially for non-technical folks and making building useful software more accessible than ever.,
But as a software developer, I still believe it’s important to understand the tech beneath the surface. Letting AI do the work without knowing why or how can be limiting in the long run.
That said, we should not shy away from using these tools, and use them to amplify our learning, speed up our workflowand push projects further. Just make sure you are still learning as you build.