AI automations · operations

Make the queue blink so you don’t have to.

I build agent tooling, retrieval pipelines, and overnight automation — with demos you can try on this page.

Scroll the work

Case studies

About

Person behind the stack

I design systems that let agents run unattended work safely: clear eligibility rules, fail-closed defaults, and logs you can audit later. I also build retrieval pipelines so writing stays grounded in sources.

I also write poems. A couple sit below — garnish, not the entrée.

Selected writing

Placeholder

I clock out in theory.
The queue does not.

Somewhere a checkbox
is the whole night’s weather.

Resumé

Derious Vaughn

Los Angeles, CA · 310-350-3717 · deriousvaughn@gmail.com

linkedin.com/in/deriousvaughn · github.com/theHaruspex

AI solutions delivery specialist who turns business problems into production automations and internal tools. Works AI-first in Claude, ChatGPT, and Cursor to design workflows, prompts, and integrations, then documents reusable playbooks so teams can adopt them. Strong in APIs, webhooks, JSON, and low-code systems (Notion). Remote-ready with stakeholder-facing delivery experience.

Skills

Claude AI Agents Prompt Engineering Workflow Automation APIs & Webhooks Notion Python Documentation

Experience

Automations and Operations Specialist

04/2025 – present

eCustomSolutions · Remote

Delivered AI-assisted automations from discovery through production: scoped workflow needs with operations stakeholders, built solutions in Claude and Cursor, and documented runbooks teammates could operate without hand-holding.
Built no-code Notion automations for an IT deployment pipeline and helpdesk intake (triage, assignment rollups, requester visibility), turning manual coordination into repeatable workflows.
Integrated systems with event-driven webhooks and REST APIs so workflow changes reached reporting in near real time instead of on a lagging scheduled sync.
Maintained living architecture documentation and sandbox prompts for workspace redesign, including MCP-based schema discovery to ground automation design in live system data.
Partnered with engineering, IT, and leadership to identify bottlenecks, capture decisions, and simplify a multi-database operations workspace.

Operations & Tech Solutions

05/2023 – 04/2025

Shades of Color · Carson, CA

Automated e-commerce and accounting integrations across WooCommerce, Amazon, and QuickBooks APIs to remove manual steps and improve reporting accuracy.
Debugged production integration failures when external systems changed and explained fixes in plain language to business owners.
Built ingestion and reporting workflows from messy sales and advertising data so leadership could rely on recurring operational outputs.

Independent engineering projects

2024 – present

Personal · Remote

Built AI-assisted outreach and scraping tools for personalized email generation, contact extraction, demographic enrichment, and structured reporting workflows.
Built an idempotent AWS Lambda ETL with Docker and GitHub Actions for batched ingestion, plus a Python SQL Server hydration toolkit for local and CSV-backed data loading.
Built a local stdio MCP server (macos-mcp) that exposes Mail, Calendar, Reminders, Notes, Messages, and Contacts to agent clients, with draft-only email so nothing sends without review.
Built a Notion schedule-driven automation trigger that parses board schedules and flips automation flags on a timer (EC2 cron + Notion API).
Built a daily environmental justice ELT joining EPA air quality with Census income data for LA County, plus a cite-first LA City Council research assistant (la-council).

Education

Coursework in Communication Studies and Honors Humanities
Azusa Pacific University · Azusa, CA

Selected experience and projects. Case studies above show the work in more detail.

Case study 01 · Ops agent

Inbound that doesn’t wait on a human to wake up.

Triage → tools → draft → pass/fail gate. Try it below.

Problem

Inbound triage is a classic ops tax: a junior person burns 30–60 minutes gathering context, drafting a reply, and second-guessing blast radius — with no structured quality gate before something looks “sendable.”

How it was solved

A LangGraph-shaped pipeline (plan → retrieve → tool calls → draft → eval) with MCP-style tools, live latency/token/$ strips, and a hard pass/fail gate. High-blast actions stay human-gated — guardrails, not autonomy theater.

Python LangGraph Tool use / function calling MCP Agent evals Observability (latency · tokens · $) Guardrails TypeScript / Node (demo UI)

Built with the stack modern ops agents use in practice: Python, LangGraph/LangChain, tool use, MCP, evals, and cost/latency observability.

ProblemSlow, uninstrumented inbound triage with no quality gate
ConstraintNo silent send; high-blast paths stay human-gated
StackLangGraph-shaped agent · MCP tools · eval scorecard · cost/latency strip
Result~2 min demo path · ~$0.10 · PASS/SKIP explicit

Interactive

Run the tower

Idle

Latency

—

Tokens

—

Est. $

—

vs human

30–60m

Pick a sample. Hit run.

Same-origin /api/triage (Control Tower logic, in-repo). Live if OPENAI_API_KEY set; else mock.

Case study 02 · Overnight automation

An opt-in that can wake a cloud agent.

Scheduled pickup with eligibility gates — not “always-on autonomy.”

Problem

Work marked “in progress” still sat idle overnight. There was no safe way to let an agent pick something up without risking double-runs, credentialed blast radius, or “it ran because a task happened to be open.”

How it was solved

A scheduled job on a home server checks a task board for a narrow eligibility rule: the task must be in progress and explicitly opted into overnight automation. Only then does it start a cloud agent with the right tools attached, clear the overnight flag (so it won’t re-fire), and write a short work log. If nothing qualifies, the job stays idle.

Python MCP Multi-agent / control plane Scheduled jobs Guardrails (eligibility gates) Task board API Cloud Agents API Home server ops

Built around MCP, multi-agent ops, Python automation, and guardrails — an agent control plane, not a classical ML training platform.

ProblemIn-progress work never auto-starts overnight
ConstraintNever start without an explicit overnight opt-in; idle when none qualify
StackScheduled check · task query · cloud agent + tools · work-log writeback
ResultOne eligible task → one run; empty set → idle

Interactive

Simulate an overnight check

Idle until a task is both in progress and opted into overnight automation.

Case study 03 · Contract literacy RAG

Ground the agent before it advises on a contract.

Retrieve clause context first → then the LLM can speak. Empty hits mean stop — no invented legal text.

Problem

A small shop is about to sign a vendor or client master services agreement. An LLM assistant that “knows contracts” will invent confident clause explanations unless it is informed first — retrieve grounded sources before it recommends, drafts, or calls tools.

How it was solved

Build a subject-scoped contracts corpus the agent queries before it acts. Embed the question, retrieve top clause snippets over a curated Nolo / CA contracts shelf, and constrain the answer to cite-only hits. If retrieval is empty, the agent stops and escalates to a human (or counsel) instead of guessing. This demo is the same retrieval step an agent would run at the edge — not a person browsing a public index sample.

Agent-grounded RAG Legal / contracts shelf Pre-advice retrieval Cite-only answers Workers AI embeddings Vectorize Pages Functions Empty-hit fail-closed

Grounded retrieval before the agent advises is the core proof — the same pattern as production RAG stacks, scoped so an LLM can explain MSA clauses from sources instead of improvising them.

ProblemLLM advice on MSAs without sources invents clause meaning
ConstraintRetrieve → cite only; empty hits → agent stops / escalates
DemoAgent retrieval step · edge embed → Vectorize → cites
ResultClause-aware hits or an honest empty set — never invented legal text

Interactive

Run the agent’s retrieval step

What an agent would pull before advising — Nolo / CA desk references via Vectorize. Not legal advice. Zero hits means stop; don’t invent a clause.