AI PRODUCTION: 5 LESSONS FROM THE TRENCHES

DEMOS ARE EASY. Production is a war. Shipping Anaqio taught me that real AI engineering happens in the plumbing, not the model.

1. PROMPT ENGINEERING IS PRODUCTION ENGINEERING

A prompt is a contract.

• ANAQIO’S PROMPT: 8 lines, 30 revisions. Encodes garment isolation, model proportions, and editorial aesthetics.
• SO WHAT: Version your prompts. Test them against benchmarks. If it’s not in Git, it’s a liability.

2. GENERATIVE ERROR HANDLING

Models don’t just crash. They fail subtly.

• THE FAILURE: Text refusals disguised as results. “I’m unable to generate this.”
• THE FIX: Explicitly catch text-only responses. Map them to user-recoverable errors. Don’t let the loading spinner spin forever.

3. BORING STATE MACHINES WIN

Simple beats complex.

• THE FLOW: pending → processing → completed/failed.
• THE STORAGE: A single Supabase row. No DAGs. No workflow engines.
• SO WHAT: Crash recovery is trivial. The DB is the source of truth.

4. SECURITY BOUNDARIES

AI doesn’t get a pass.

• VALIDATE EVERYTHING: Path ownership (!path.startsWith(userId)), URL safety (SSRF allowlist), and Provider locking.
• SO WHAT: Treat AI endpoints like file uploads. Assume malicious intent.

5. AUTHENTICITY > OVERPROMISE

Ship the simplest architecture that works.

• THE BET: Single API call (Gemini) > 5-stage custom pipeline.
• THE RESULT: Shipped in weeks, not months. Real customers, real revenue.

THE TAKEAWAY

Ship the simplest path. Optimize when volume dictates. Boring engineering wins.