The original headline is: "I Scored 10,000 Job Listings a Day With GPT-4. Here's What Broke"

Short summary

A developer built a production system scoring 10,000+ job listings daily with GPT-4, learning critical lessons about moving from demo to production. Key findings: OpenAI's Batch API reduces costs by 63%, multi-chunk extraction with focused schemas improves accuracy from 12% to 2% error rate, and token bucket rate limiting prevents pipeline degradation. The unresolved challenge: scaling AI description rewrites profitably, with DeepSeek emerging as a potential 23x cheaper alternative.

• •Batch processing reduces cost 63% ($86→$32/day) at the cost of latency
• •Multi-chunk extraction with explicit schemas cut errors from 12% to 2%
• •Token bucket rate limiting with 100ms buffer prevents cascading 429 errors
• •Presence guards in schemas prevent hallucination on optional fields

Generated with AI, which can make mistakes.