Work
Four pieces of work, each with the problem, what I built, and what changed.
A production agent for datacenter network investigation
Problem. On-call engineers spent about an hour gathering device state across a large AI datacenter core network before they could even start reasoning about an incident.
What I built. A terminal agent that operators question in plain English. It fans out across the fleet through typed tools and lands on conclusions the evidence supports. I wrote a custom agent loop instead of adapting a general coding agent, so safety lives in code and tests: structurally read-only, operator approval on any change, and deterministic investigation paths (same question, same fleet state, same answer). Every device query returns structured, device-attributed results, and saved output is checked line by line against real tool output before it's written.
How I know it works. Three layers of evals on recorded, replayable runs: an offline regression suite that gates CI, a live grounding eval that measures fabrication rate, and a cross-model behaviour baseline. Regressions block shipping.
Result. Incident investigation went from ~1 hour of manual state-gathering to minutes, and new engineers use it to learn the network. Write-up: The Model Is the Easy Part.
Zero-touch provisioning, three times
Campus/datacenter. ZTP/POAP for Catalyst 9k and Nexus 3k/9k: 96% faster provisioning.
Edge. A ZTP pipeline for field edge devices: 75% less operator-involved configuration, plus automated firmware upgrades (10× operator efficiency) and firmware QA for 4G/5G edge routers.
Validation. Pre-deployment checks that prove a device is ready before it's handed to production, on top of a single-source-of-truth network model (IPAM/CMDB).
CMDB remediation automation
Automated remediation that keeps the CMDB aligned with the real network, instead of people reconciling it by hand: 98%+ less operator involvement.
Network automation in the field
Five years as principal solutions architect for a network automation platform: led automation proofs-of-concept for large enterprises and public-sector networks, built repeatable demo environments with Docker and templated config automation, and kept a ~90% technical win rate. Earlier: sales engineering for big-data and embedded-Linux companies, and co-founder/technical lead of an open-source storage-replication company.