A concrete story about business-driven engineering trade-offs: a vibe-coded service with a known production bug he deliberately leaves unfixed.

The code execution service

  • He was paying $3,000/month for a third-party code execution service (runs user-submitted code on NeetCode) 32:02
  • With AI-assisted coding, he estimated a rewrite would take a couple of weeks — it actually took 2–3 days (though "it did take coding skills. Like if I didn't know how to code, I would not have been able to do it") 32:25
  • Self-hosting now costs ~$200/month vs $3,000 32:47

The bug and the trade-off

  • The service has a bug (likely a memory leak): every couple of days one or two instances crash 32:47
  • Root-causing it would probably take longer than the 3 days it took to build, so instead he runs ~4 redundant instances; crashed ones are replaced within a couple of minutes 33:14, 38:17
  • "The engineer in me hates that… but the business value makes no difference. There has been practically zero outages. I have less outages than LeetCode" 36:05
  • He rolled it out carefully: small traffic percentage for a couple of weeks against the original service, expecting issues that never materialized 37:52
  • Fixing it would only save "a couple hundred bucks more" in server pool size — not worth it. The new service is also faster (better hardware) and no user has complained 37:33, 38:17

The host's pushback

  • Gergely challenges: isn't this typical AI-coding behavior — replacing a SaaS with something subpar that's now too hard to fix because you didn't write it? 37:12
  • Resolution: packaged together (cheaper + faster + replicated), it's overall better than before — an obvious business decision that raises the question of how perfectionist to be when a problem is already solved well enough 38:41
  • Navi concedes it still bothers him, but he's learned to prioritize actual value, short and long term 39:03