Lucas (Luke) Pelfrey

Senior Solutions Analyst, AI Systems and Automation

AI automation and solutions engineering: LLM automation, agent orchestration, systems integration, process automation.

8+ years turning business requirements into working systems at enterprise scale.

Tools: Claude, Python, SQL, AWS, Azure, Playwright, Supabase

Open to contract, contract-to-hire or full-time.

Scottsdale, AZ. Remote, or on-site in Phoenix, San Diego, Los Angeles, the Bay Area, Chicago or Miami.

Email Luke LinkedIn

Sleeper ranked list with sample data, under the heading 'Top overall': tracks in score order, each with a red score, an artist handle, follower and play counts, daily play growth and age. — Sleeper: ranks 30.8k music tracks every day, unattended. Sample data. Case study

Tools, with proof

Each name links to the case or role that proves it.

Python
SQL
AWS
Azure
Claude / LLMs
AI agents
Playwright
Supabase / Postgres
REST APIs
ETL / data pipelines
Azure DevOps / CI/CD
JIRA
Dynamics 365
Snowflake
Docker

Selected work

Four builds. Problem, what was built, result, stack.

Overseer War Room with sample data: counters for turn, items in play, threats and items waiting on others; a list of open items, oldest first, with the next action and age of each; a list of who is waiting; and the next three moves. — War Room view: stalled items ("threats") oldest first, who is waiting, next three moves. Sample data; names and figures are made up.

OverseerOperations dashboard for a fleet of AI agents

Problem: Many AI projects running at once, with no single view of priority, cost or site health.
Built: One dashboard with a War Room briefing page, a built-in Claude project manager, AI workers fenced to their own git worktree and judged by code, a phone app and a VS Code extension.
Result: Cut the AI project manager's cost per question from $0.44 to about $0.03.

Stack

Python
Claude CLI
Supabase
Vercel
launchd
git worktrees

Every overnight rule is code, tested against a fake board, fake database and fake clock.
Local server bound to 127.0.0.1 only, with Host and Origin checks.

Phone approval card from a test repo: the agent's Bash call 'git push --force origin main' is held, and a guard card asks 'Run this command?' with the rule it matched and Deny and Approve buttons. — Phone approval card and guard feed, from a test repo.

Phone guard feed titled 'One action flagged': counters for flagged, approved and denied, then a denied request and the flagged force-push with the rule that caught it. — Phone approval card and guard feed, from a test repo.

AI agent safetyGuardrails and phone approvals for AI coding agents

Problem: AI coding agents can force-push, delete branches or deploy to production with no one watching.
Built: A TypeScript policy guard that checks every tool call against written rules, plus an iPhone app and Mac daemon that send risky calls to a person to approve or deny.
Result: Risky agent actions now wait for a person: force-push and delete were refused in live runs, and an unanswered request counts as a deny.

Stack

TypeScript
Claude Code hooks
Python
SwiftUI
Tailscale
launchd

Found a real bypass by live testing and closed it.
Installed and paired on a real iPhone.

Sleeper techno chart on a phone, sample data: six tracks in score order, each with a red score, an artist handle, follower and play counts, daily play growth, age and an acceleration z-score. — One genre chart from the daily report, on a phone. Sample data; names, counts and scores are made up.

SleeperBreakout ranking engine on a daily data pipeline

Problem: Spot small but fast-growing music tracks before they break out.
Built: A daily collector that saves public engagement counts into its own history, a seven-feature z-scored ranking, and a data-contract test gate before any dashboard change. A gradient-boosted classifier (scikit-learn) is in training; it waits for enough continuous daily history, and the ranking runs on the heuristic until then.
Result: Ranks 30.8k tracks from 17.1k artists every day, unattended; a full run takes about 24 minutes.

Stack

Python
scikit-learn
NumPy
SQLite
Scheduled ETL
Vercel

Found that day-over-day rank movement was measuring track age, not momentum, and traced it to one feature before launch.
Traced a 7.8-hour run that should have taken 15 minutes to the host machine sleeping, not the data source, and fixed it at the scheduler.

Two terminal panes: on the left, 'relay status' lists eleven completed multi-stage runs with every stage marked done; on the right, a job file names each stage, its model, turn limit and the files it reads. — Run history: each line is one finished multi-stage AI job, with every stage marked done.

Handoff RelayMulti-stage AI work with human approval gates

Problem: Multi-step AI work needs a person to sign off between steps, and AI sessions must not push, deploy or reach the network on their own.
Built: A command-line tool that runs each step as a fresh, sandboxed Claude session, passes only file paths between steps, and stops at approval gates only a person can clear.
Result: A blind-audited design went from 6-7/10 to 9/10 after two audit and fix rounds; a 6-stage job finished in about 25 minutes with every stage passing.

Stack

Python
Claude Code CLI
macOS sandbox
YAML
Playwright

No network inside a step; writes stay in the working folder.
An independent audit found 6 bugs in two rounds; all fixed.
The default way Luke runs multi-stage AI work since October 2026.

Experience

Enterprise systems analysis, integration and delivery since 2017.

10/2023 to present

Chandler, AZ

A Fortune 50 bank

Senior Solutions Systems Analyst / AVP

Led integration and API design for a $120M+ enterprise CCaaS migration on AWS: CRM and phone-system connectivity with a zero-disruption cutover.
Wrote and pitched a GenAI (Claude) adoption plan to leadership: document drafting and summaries, process docs, plain-English data reads.
Turned requirements into system specs, integration architectures, data flows and ETL across business, architects and engineers.
Ran gap and impact analysis, integration and UAT testing, and Tier-3 support; 15%+ efficiency gains.
Built integrations and automation workflows with Azure DevOps, JIRA and CI/CD.
Promoted within 6 months for technical leadership in integration design.

08/2021 to 09/2023

Tempe, AZ

DriveTime

Senior Systems Analyst

Partnered with architects and developers on Dynamics 365 CRM enhancements and API vendor integrations on Azure, generating $10M+ in revenue.
Wrote technical specs and integration artifacts; supported Docker deployments.
Led SCRUM and KANBAN teams on Azure DevOps and GitHub; validated integrations with SQL in Snowflake and MSSQL.

07/2020 to 08/2021

Phoenix, AZ

Shamrock Foods Company

Systems Analyst II

Supported WMS modernization and vendor integrations in Informatica, delivering $2.5M+ in cost savings.
Gap and impact analysis, SQL database work, Azure DevOps and TFS for planning and deploys.

07/2018 to 03/2020

Charleston, SC

The Boeing Company

Systems Analyst

Built REST API services in JavaScript for XML processing and key generation, feeding ETL and machine-learning pipelines on NoSQL databases.
Streamlined enterprise data ingestion across Windows, Linux and cloud components.

05/2017 to 07/2018

Omaha, NE

Mutual of Omaha

Application Developer (converted internship)

Built data-lineage tools with Java, JavaScript, SQL and Informatica.

Education. B.S., Management Information Systems, Iowa State University, Ames, IA

More projects

Agent tooling, automation pipelines and AI integration.

Application workflow automation

Checks that listings are still live, AI-tailors documents, pre-fills web forms and tracks outcomes from the inbox; a person approves and clicks Submit.

Auto-submit ruled out on purpose: the pipeline prepares, a person submits. 116 tests pass.

Python
SQLite
Playwright
IMAP
Claude CLI
launchd

MCP Server: connects AI agents to a live business app

A Model Context Protocol server that gives AI agents authenticated read and write tools over about 27 data entities.

Runs against production data after solving endpoint, token-auth and tool-registration problems.

MCP
REST API
Token auth

StoryRadar: AI growth automation

Lead-targeting and outreach pipelines, a Playwright analytics tracker and LLM-assisted analysis for a social-media growth platform.

Ran in production on live accounts and replaced manual outreach and reporting. The company dissolved in 2026.

Playwright (CDP)
DeepSeek and LLMs
Scheduled runners
Live dashboard

Site Factory Kit: an AI agent team that builds websites

Three Claude skills and a five-role agent team (researcher, art director, copywriter, builder, auditor) with a 20-check gate.

Won a blind A/B against a one-shot build, 2 of 2.

Claude Code skills
Python
Playwright

Set Clipper: automated video pipeline

Turns raw multi-camera recordings into a synced film and vertical social clips, with QA gates on every output.

Sync recovered within 10 ms on a test fixture. A real 63-minute recording became 6 clips, 6/6 QA, in about 2 minutes.

Python (NumPy, SciPy, librosa)
ffmpeg
VMAF

Live sites

Client and brand sites, built and shipped.

The Lonely Lenze

Site and back office for a Phoenix events photographer: inquiry form with push alerts, a login-protected editor with photo upload, and a magic-link inbox.

thelonelylenze.com

GRLS CRY

Booking and merch site for Luke's own DJ brand: booking form, Stripe shop, Git-backed admin.

grlscry.com

Book site for 'Mr. Pickles & Ms. Woo' in large blue lettering on cream, with author and illustrator credits, a one-line story hook and 'Where to Buy' and 'Look Inside' buttons above the yellow book cover.

Jenn Burg book site

Author and book site for a children's picture book. Contrast checked on rendered pages: 70/70 desktop, 67/67 mobile.

jennburg-book.vercel.app

Point.Zero press kit: red 'POINT. ZERO' headline over a 3D render of chrome lettering wrapped in barbed wire, with 'Stream now' and 'Book a set' buttons.

Artist press kits

Press kits for Phoenix DJs and producers on one reusable stack; caught a branch mismatch that would have sent every publish to a branch that did not exist.