For me, 2025 was about one main thing: using AI to accelerate the development efficiency of the Databend team.

AI

I built testing and support systems with AI, using them to locate customer issues faster.

I tried many models, but ended up with just two on my desktop:

  • Codex (gpt-5.2)
  • Claude Code (Opus 4.5)

They handled most of the heavy lifting: analyzing logs, reproducing bugs, writing glue code, and fixing issues.

Yes — all the systems listed below were built with AI assistance.

The Unsolved Problem: Determinism

One problem remains unsolved: testability and correctness.

When a system undergoes multiple rounds of AI-driven modifications, each round adds a bit of uncertainty. After enough rounds, while delivery speed increases, trust decreases.

How do we maximize determinism? I still don’t have a good answer. Hopefully 2026 will bring tools to solve this.

Systems Built in 2025 (Vibe Engineering)

So what did I build this year?

Top1: Databend Testing System

The Databend testing system: comparing performance and functionality against other top-tier data warehouses to catch regressions early.

Databend testing system

Top2: databend.com

The official website. Fast, clear, real-time updates.

databend.com

Top3: Public S3 Bucket s3.databend.cloud

A public bucket for sharing build artifacts and files.

s3.databend.cloud

Top4: Unified Key Management for Codex / Claude Code / Gemini

Unified account and key management: one API key, multiple providers, multiple upstream keys for failover. Greatly simplified integration and daily use.

Unified AI key management

2026

My biggest wish remains unchanged: bring more determinism to AI-assisted engineering.

If we can systematically solve correctness, testability, and regression detection, the speed ceiling can be raised another level.