I am the kind of person who learns by doing. When I watch a video course, I feel that I grasp about 80-90% of the material. However, when I start working on a project, I often find myself wondering how things were done, even if I watched the course just 2-3 days ago. It is crucial to apply knowledge to a project that is entirely different from what the instructor demonstrated; this approach leads to a deeper understanding.
I decided to apply my knowledge from the course to Cheapta, a project designed to help users find cheap flight tickets. My first step was to create a chatbot for Cheapta. Initially, I developed a voice agent using n8n and ElevenLabs. To reduce costs by $20/month, I opted to self-host n8n on my server. I then consulted Claude Opus 4.6 to create a manual implementation plan, which I followed to build the n8n workflow manually.
The voice agent I created with ElevenLabs sends requests to the n8n webhook. However, I ran out of ElevenLabs' free credits while testing it. As a result, I decided to switch from a voice-based chat to a text-based chat.
I first experimented with OpenAI GPT-4, which cost $0.50 for 1,182,000 tokens (approximately 23,640 tokens per cent). I then switched to DeepSeek Flash that charged $0.01-$0.02 for 292,000 tokens (about 146,000 tokens per cent). I found that DeepSeek Flash performed better than GPT-4. A person I met in a coffee shop, who was working on his own project, recommended DeepSeek because it is much more cost-effective. Initially, I had dismissed DeepSeek without trying it, but I now understand that it is a better option in terms of value.
I have successfully implemented an AI chat assistant on cheapta.com. You can try it out to find cheap flight tickets!