rescue culture trap

We hired an AI engineer. Smart, solid research background, could talk about embeddings and fine-tuning all day. Role was supposed to be 80% backend, 20% AI. Didn’t work out.

The pattern started small. They’d hit a production bug, a 429 rate limit error, a failing deployment, a flaky test, and instead of debugging it they’d escalate. Not “I’ve tried X, Y, Z and I’m stuck.” More like “this is broken, can someone look at it?”

And I’d look at it. Every time. Because it was faster, we were shipping, and I told myself I was unblocking them.

What I was actually doing was building a rescue loop. Every time I solved their problem instead of coaching them through it, I reinforced that the correct response to a production issue is to wait for someone else to fix it. Their debugging skills atrophied because they never had to use them.

The root causes were visible in the interview if I’d known what to look for:

Role drift. We needed someone who could own production backend work and occasionally apply AI techniques. We hired someone whose identity was “AI researcher” and expected them to care about deployment pipelines and error handling. The 80/20 split was always going to create friction.

Standards erosion. PRs that should’ve been 200 lines were 800. Error handling was optimistic. Tests covered the happy path. I’d give feedback, they’d fix the specific issue, but the underlying approach didn’t change. I kept thinking “they’ll level up” instead of recognizing a pattern.

Over-shielding. This one’s on me entirely. By solving their problems I prevented them from developing the production mindset the role required.

What I screen for now:

• “Tell me about a time you shipped ML to production and it broke.” If they can’t tell a debugging war story with specific details (the error, what they tried, how they traced it), red flag.
• “Walk me through debugging a 429 rate limit error.” Not theoretical. Actual steps.
• “What’s your PR size limit?” No right number, but the answer reveals whether they think about reviewability.
• “If you’re stuck for 2+ hours, what do you do?” Red flag: “Ask my manager.” Green flag: “Check logs, reproduce locally, read the source, then ask with context.”

Hardest part was accepting my role in it. The hire didn’t fail because they were bad. They failed because I created an environment where they didn’t need to be good at the parts of the job that mattered most. Rescue culture feels like support. It’s actually neglect.

We’ve since shifted from “AI production engineer” to “SaaS engineer.” The AI work is now led by someone who actually wants to own it end-to-end. Better fit, better outcomes.