AI
I've deployed AI across three distinct operational modes. Each requires different thinking about where humans belong in the system—and different metrics for whether it's working.
Automation
Removing humans from loops where they add latency but not judgment. The hard part isn't the technology—it's knowing which loops qualify.
Document Processing Pipeline
A 40-person ops team manually reviewed and classified 3,000+ documents daily, with a 72-hour processing SLA that clients were pushing back on.
What I did: Designed and implemented an ML-based classification and extraction pipeline, defining the confidence thresholds that determined when documents could be auto-processed versus flagged for human review.
Result: Reduced average processing time from 72 hours to 4 hours. Human review dropped to 12% of documents. Error rate improved from 4.2% to 1.1%.
Customer Support Triage System
Support volume had tripled in six months. The team was spending 60% of their time on routing and categorization rather than resolution.
What I did: Built a triage layer that classified incoming tickets by intent, urgency, and required expertise, routing automatically to the right specialist queue with relevant context pre-loaded.
Result: First-response time dropped 65%. Specialist utilization improved by 40%. CSAT increased 18 points within one quarter.
Augmentation
Making humans better inside the loop. The design challenge is surfacing the right information at the right moment without creating dependency.
Decision Support for Account Managers
Account managers were making renewal and expansion decisions based on intuition and incomplete CRM data. Win rates on expansion deals had plateaued at 31%.
What I did: Designed an AI layer that synthesized product usage data, support history, and market signals into actionable briefs delivered before every client touchpoint.
Result: Expansion win rate increased to 47%. Average deal cycle shortened by 3 weeks. AMs reported spending 40% less time on pre-meeting research.
Engineering Code Review Assistant
Code review bottleneck was adding 2-3 days to every PR. Senior engineers were spending 30% of their time reviewing junior work for patterns they'd already documented.
What I did: Implemented an AI assistant that pre-reviewed PRs against team style guides and architectural patterns, flagging issues and suggesting fixes before human review.
Result: Review cycle time dropped from 2.5 days to 8 hours. Senior engineer review time reduced by 55%. Critical bug detection in review improved 23%.
Agentic Systems
Creating entirely new operational loops that didn't exist before. This is the frontier—and where most organizations are making their most expensive mistakes.
Autonomous Vendor Negotiation Agent
Procurement was a manual, relationship-heavy process. The team handled 200+ vendor contracts annually with limited leverage on tail-spend categories.
What I did: Designed an agentic system that autonomously managed initial negotiation rounds for routine procurement, including market price benchmarking, term comparison, and counter-proposal generation.
Result: Achieved 12% average cost reduction on tail-spend categories. Procurement team capacity freed to focus on strategic vendor relationships. Time-to-contract reduced by 60%.
Multi-Agent Incident Response Coordinator
Production incidents required coordination across 4-6 teams. Mean time to resolution averaged 4.2 hours, with most time spent on diagnosis and coordination rather than fixing.
What I did: Built a multi-agent system where specialized agents handled log analysis, dependency mapping, runbook execution, and stakeholder communication in parallel, with a coordinator agent managing the workflow.
Result: MTTR dropped to 1.4 hours. Human intervention reduced to final approval and edge cases. Post-incident review quality improved as agents maintained detailed decision logs.