John Zolton

• Launched new Submittal Register product by productionizing NLP research into a Prefect-orchestrated AWS pipeline, delivering an app-ready service end to end
• Led cross-team integration by defining data contracts and coordinating API interfaces between NLP and frontend/app teams, unblocking delivery and aligning pipeline outputs with product requirements
• Drove major latency improvements through parallel execution and pipeline tuning, processing ~3,000-page documents in ~5–8 minutes (~4× faster than early runs, beating <10-minute target)
• Implemented token-efficient consolidation and deduplication across parallelized paths, reducing redundant LLM work while preserving deterministic outputs
• Improved extraction quality through evaluation-driven iteration, achieving ~99.6% IoU on page-range detection and higher section-splitting and data-extraction accuracy
• Optimized source-text highlighting by constraining bounding-box search to predicted pages with ±2-page fallback (capturing ~99.8% of misses), reducing local runtime from 10–15 minutes to ~4–5 minutes while maintaining highlight coverage
• Added production robustness for LLM-dependent steps with retry handling and token-range chunking to prevent request failures and improve pipeline reliability
• Supported Submittal Review v3 stabilization on TRACI, helping reduce failed-run rates from ~20–30% to ≤1% through production hardening and targeted fixes
• Refactored agent orchestration patterns for improved clarity, maintainability, and reliability
• Provided critical bug-fix coverage during senior engineer PTO, keeping releases on track

Frederick, MD