James Kirkman

CCE SRE Mission Control, agent-ready knowledge platform (founder and primary engineer)

• Founded the team’s source of truth, used every day by the whole ~15-person team for nearly everything: handbook, knowledge vault, and incident records as reviewed Markdown in git. Built from an empty scaffold in ~4.5 months.
• Split the knowledge into two layers: a vault structured for LLMs (196 entries) and a wiki written for people (160+ Confluence pages), kept in sync by a round-trip converter and publish-on-push pipeline with drift detection.
• Designed a four-stage knowledge pipeline: capture from team sources such as Slack, Zoom, and Confluence; an LLM proposes edits; a person reviews them; then publish. Because every fact traces to its source, the vault answers questions nobody planned for, like why a set of network devices lacked instrumentation, traced to a month-old meeting recording.
• Built a context layer that lets any agent, in the IDE or in Slack, find the right knowledge without flooding its context: per-repository AGENTS.md files, a context-budget audit that blocks pushes over budget, and 14 rules and 12 skills that gate agent-created tickets and publishing.

CCE-SRE-Expert, the platform’s production Slack agent (architect and primary engineer)

• Built the agent in Python (Slack Bolt) on Claude via Databricks Model Serving. Internal customers use it almost daily, and over 60 days it correctly helped 82% of them without an SRE stepping in, often by debugging their code rather than pointing them to docs.
• Gave it 20+ tools, including live lookups on the company’s cloud-policy MCP server over OAuth client credentials, and a human-in-the-loop write path so teammates can update CCE SRE Mission Control from Slack: the agent files a Jira ticket, stages edits, shows a diff, and opens a draft PR on “ship it.”
• Engineered retrieval as an agentic tool loop over vector search (BGE-Large in Databricks Vector Search), file reads, grep, and directory listing, replacing single-shot RAG after eval failure analysis showed the bot couldn’t re-query, follow links, or read whole documents.
• Put safety in the harness: write tools the model can’t see outside one approved channel, draft-PR-only writes, and post-answer verifiers for rules prompting couldn’t hold. Every production miss becomes an eval case.

Agent architecture and multi-agent strategy

• Wrote the team’s AI architecture roadmap: a reversible path from one agent to per-domain experts, with each step gated by a measurable trigger. We’ve held at step one: a cost model put self-built domain experts behind a router at ~12× the tokens, and evals showed one agent with the right facts in the vault was enough.
• Evaluated an enterprise agent platform as build vs. buy: it could replace the agent’s runtime but not the knowledge beneath it, so I recommended an eval-gated side-by-side and publishing the vault over MCP either way.

AI-native engineering practice

• Designed the git-worktree workflow my 15-person team adopted for large or conflicting parallel agent sessions, offering a worktree only where two sessions could commit over each other.
• Guarded what agents write: rules, skills, and a sensitive-data scanner keep private IP addresses and raw data exports out of the repository, enforced at pre-push, at publish, and in CI.
• Re-based sprint estimation on human–LLM pairing, right-sizing an overcommitted 49-point sprint to about 7.

Reliability engineering

• Lead Tier-0 network observability for hundreds of critical devices ahead of the 2026 holiday season, from a per-device gap register to phased alerting. A log audit found a delayed feed had left existing alerts nearly blind, so feed-health alerts come first, and an alert counts as done only after an end-to-end fire test.
• Handle on-call PagerDuty triage, write blameless RCAs for high-severity incidents, and set up the team’s RCA library.

Hive On Demand, self-service distributed load testing (designer and owner)

• Designed and built Nike’s self-service load-testing platform end to end, so teams could build ad hoc and experimental load tests without writing code: React SPA, Lambda/API Gateway controller, containerized Gatling runner on ECS, and S3 test data and reporting, plus scenario chaining, response validation, and custom load injectors.
• Launched it in 2021 and ran it on AWS for 5+ years (CloudFormation, CloudWatch, CloudFront, DynamoDB, IAM), through an account migration, an AWS SDK v3 upgrade, and 2026 routing to next-generation load generators.

Performance Reporting Portal (lead developer)

• Led development of the portal (React, Vite, D3): dashboards that turned real-time web and mobile data into metrics the business acts on, from Largest Contentful Paint to bounce rate and conversions, for Nike.com and the Nike apps.
• Helped build the web data pipeline (New Relic, AWS Glue, PySpark) and led the mobile real-user monitoring pipelines (Lambda, API Gateway, OpenSearch) for iOS and Android, Global and China, with automated trend reports.
• Took CrUX-based benchmarking from MVP to production and used it to build a competitor dashboard comparing Nike’s web performance with its competitors’ across phone and desktop.

AI enablement (2025–2026)

• Led the team’s adoption of AI coding agents: AGENTS.md context across 13+ repos, a shared rules-and-skills repo, and Cursor domain agents for the Hive stack and mobile performance, used to harden Hive On Demand.

Contractor through NewCoe; converted to a full-time Nike employee in 2017.

• Ran daily performance tests on individual microservices with Hive, the work that delivered the most value: they found bottlenecks fast, exposed architectures that didn’t scale, showed how gracefully services failed, and gave teams upper and lower bounds for their systems.
• Ran launch-scale load tests for Nike’s high-demand launch platform: production traffic modelled as user-journey buckets in Gatling, load shaped per launch type, and fleets of synchronized cloud load generators driven past security and bot detection. After a difficult launch the year before, a full-scale “dark prod” test predicted the next launch, which had no major outages through the holiday season.
• Built a functional and performance test framework for the User Lifecycle Services platform, with mocks that could be slowed or failed on demand, exposing a 100 rps Step Functions throttle that blocked the subscription launch.
• Ran launch-readiness performance assessments for consumer platforms such as WeChat Mini Program, Cart, Orders, and identity, with load models built from production traffic (Splunk, New Relic).

• Built web application features and the automated unit, integration, and functional test pipeline; led testing strategy.

• Led front-end development on AAA’s Sixth Sense web application as technical lead.
• Wrote a Java service that ingests third-party data files and routes them to brokers by configured rules.

• Programmed Campbell Scientific data loggers to measure water-quality metrics, and installed them in the field at spring-water sites across rugged Northern California terrain.