Reliable tool-using AI for network operations
Catalyst Center's AI assistant answers over live inventory, health, issue, and
troubleshooting data — where a plausible-sounding answer can still use stale context,
pick the wrong capability, omit a filter, or stop after one page of results. I moved
that behavior into owner-correct boundaries: typed tool schemas, current-turn
grounding, query builders, deterministic response shaping, and executable contracts,
with fixes delivered across six workflow families.
AI evaluation as a release control system
I own the evaluation and promotion gates for the shipped assistant: evaluations
run daily, before my merges, and before twice-weekly promotions — and they catch
regressions every week. Failures are traced through agent selection, tool calls,
arguments, raw results, and delivery status rather than trusting an aggregate score.
The offline replay platform I built is used by developers and testing teams and has
supported resolving multiple customer-reported defects.
Scaling network analytics without weakening correctness
Network health APIs repeat expensive graph and metric queries under concurrent
dashboard traffic. I delivered analytics APIs with query-level sorting and pagination,
asynchronous execution, single-flight work sharing, and selective caching keyed by
authorization-safe request identity — validated under concurrent load with every
response equivalent to the uncached path.
Owning Path Trace, a distributed path-computation service
Sole owner of the service that reconstructs how traffic moves through switched,
spanning-tree, equal-cost, routed, and policy-controlled segments — on a platform
whose largest documented profile supports up to 18,000 devices and 1.5 million ports.
Architecture work spans site authorization, overlapping address spaces, IPv6-aware
paths, on-demand ACL collection, and calculation-scoped spanning-tree state.
Power telemetry, device to dashboard
An end-to-end energy analytics capability across four repositories: device
telemetry parsing, streaming aggregation, graph-backed APIs, dashboards, and automated
acceptance coverage. Allocation, consumption, and savings stay separate first-class
metrics, with units, null behavior, time granularity, and late data carried through
every layer.
A resilient device-security state lifecycle
Designed and shipped a dual-source lifecycle that detects when managed devices
enter or leave a less-secure operating state — choosing redundant telemetry over a
simpler single source to avoid missed transitions, then documenting and instrumenting
the races that choice buys. The resulting security-state field is part of the public
Catalyst Center API surface.
Keeping issues and automated diagnosis trustworthy
Active operational issues could silently disappear at retention expiry while the
underlying condition persisted. I built the scheduled lifecycle that verifies the
resource still exists and republishes unresolved issues, and fixed automated
root-cause workflows that reached false conclusions — a wrong source of truth here, a
reused execution scope there. Encoding expert troubleshooting as repeatable inference
is the product.
Reproducible incident forensics
Led AI incidents that crossed orchestration, model, tool, and transport
boundaries: provider throttling hidden by graceful fallbacks, successful root traces
hiding failed child calls, completed graphs that never delivered a terminal event. I
converted the recurring work into reusable methods and a secure trace exporter —
paginate to exhaustion, verify integrity, protect credentials — now used in daily
customer-environment triage.