The Trigger: A Broken Kindergarten App

As a member of the kindergarten parents’ committee, I attended a committee meeting yesterday where one complaint kept resurfacing: the kindergarten’s app is genuinely bad.
What is presented as an “app” is essentially a web page wrapped inside a mobile shell. It is slow, unresponsive, and media content—especially photos and videos—often fails to load.

After the meeting, I outlined a product concept and basic feature plan and reached out to a software outsourcing team in China. My initial approach is intentionally minimal:
👉 design a single main-screen UI to validate visual quality and interaction flow before deciding whether the idea is worth deeper investment.

This issue points to a clear market gap.
Kindergartens in Germany are undergoing digitalization, yet most existing solutions are outdated and built with a “just usable” mindset rather than a user-centric one. I see this as a meaningful opportunity worth exploring.

---

Research Progress Update

Today’s work focused on model evaluation and inference infrastructure.

I began by deploying a 16k-context model on an RTX 6000 Ada, but soon ran into a subtle networking issue.
Because the RunPod instance was hosted in the US and I was using an HTTP-based API, the Cloudflare timeout settings conflicted with my local and evaluation-side timeouts, causing the inference to stall indefinitely at step 1.

After identifying the root cause, I switched to an A100 setup.
Inference speed improved dramatically, and the overall pipeline behaved much more reliably, with noticeably better submission behavior.

---

Issues and Findings

After adjusting the YAML configuration, the run failed halfway due to a simple but costly oversight:
Docker image pulls require authentication, and I had forgotten to log in.

I restarted the process and documented this step carefully to avoid repeating the mistake. However, the more serious problem emerged during result analysis:

• The vast majority of submitted patches were empty
• The few non-empty patches were clearly hallucinated
• Almost none were practically applicable

This led to an unavoidable conclusion:
👉 Qwen2.5 Coder 3B likely lacks the fundamental capability to handle this class of tasks.
Using this model as a baseline for such evaluations is, in practice, unrealistic.

---

Next Steps

The outcome is disappointing, but the direction is now clearer.
I need to search for a more viable approach—one that balances model capacity, compute cost, and practical effectiveness.

The search continues.
The road is still ahead.