AI-native rewrite cut a 9-month rebuild to 4 months, then broke QA
alex_verem · x · 2026-07-29
A case study argues that “AI-native engineering” often sells speed while hiding the QA cost.
The project involved a PE-backed lending platform with an 11-year-old Java monolith and a large amount of regulatory logic. One bad origination decision is not a normal bug there — it is a compliance event.
What they changed first
- No code in week one
- Scanned the repo and built a knowledge graph
- Produced about 340 markdown files covering modules, dependencies, data flow, and domain terms
- Kept one doc focused on a narrow domain area instead of spreading context everywhere
The author says the team cut a 9-month rebuild to 4 months with 2 engineers and 147 merged PRs in the first 90 days, but then broke QA. The point of the write-up is that speed is only one part of the story; reliability and regulatory correctness are the harder part.
More from coding & agent
- Replication finds agent experience distillation preserves 44.1% of ICL gains on SWE tasks — burny_tech · 2026-07-29
- Agensis turns AI agents into team members with local inference and free agent seats — jasonkneen · 2026-07-29
- Grok Build adds MCP CLI controls, instant cold start and a /delete session command — mark_k · 2026-07-29
- LLM coding workflow adds a review gate to stop every retrieved method from being implemented — hypergraphr · 2026-07-29
- Yangqing Jia and Andrew Ng Announce AI Startups: Engineering Agents and Personalized Education — 机器之心 · 2026-07-29
- A GPT code-fix screenshot turns into a classic overconfident AI meme — hrishioa · 2026-07-29