For the last three months, I have been on maternity leave and preparing for a major change in my life.
I also participated in more than ten hackathons.
I did not plan to enter so many. I wanted to keep learning, build a few ideas, and use this time to challenge myself. Then one project became another, one deadline followed the next, and suddenly I had a collection of things I never expected to finish.
I worked in whatever windows I had. Some sessions produced a complete feature. On other days, progress meant fixing one bug, improving one screen, or finally understanding why an API call was failing.
It was not a perfect routine. It was simply the routine that worked for me.
My latest project: QuoteX
For the
Global AI Hackathon with Qwen Cloud
, I built QuoteX, an AI agent for cross-border commerce.
I would appreciate your feedback
I wrote a separate technical article explaining QuoteX’s architecture, mistakes, tool-calling design, and evaluation:
I Gave Qwen Six Tools, but Not the Right to Set a Price
A buyer can describe a product using text, voice, or an image. Qwen understands the request, plans the workflow, and selects the required tools.
But I made one important architectural decision:
The AI can reason, but it cannot invent a price.
Pricing, policy checks, and commercial calculations are handled by deterministic services. A human must approve the final quotation before it can be sent.
I tested this architecture with 42 adversarial scenarios. The governed system passed all 42, while the direct AI baseline passed 28.
What I learned
Participating in so many hackathons taught me more than I expected.
I learned that building a working AI demo is not the same as building a trustworthy product. I learned that presentation matters, but reliability matters more. I also learned that asking for feedback can be more difficult than writing the code itself.
This experience reminded me that progress does not always happen under perfect conditions. Sometimes it happens in short, tired sessions between ordinary responsibilities.
I would genuinely appreciate your honest feedback:
- Is the boundary between AI reasoning and business authority clear?
- Would you trust this architecture in a real commercial workflow?
- What should I improve next?
Thank you for reading a small part of my journey.
