Derious Vaughn

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.

Resume

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.

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

Contract literacy for a small shop.

Subject-scoped legal corpus → retrieve → cite-only. Empty hits mean stop and escalate.

Problem

A small shop is about to sign a vendor or client master services agreement. People on the team do not know what indemnity, IP assignment, or liability caps actually do — so they either rubber-stamp the PDF or invent confident-sounding explanations.

How it was solved

Build a subject-scoped contracts corpus, retrieve relevant clause explanations, and answer cite-only. If retrieval is empty, stop and escalate to a human (or counsel) instead of guessing. This page’s interactive demo uses a public sample with vector search at the edge (embed the query, retrieve top snippets, cite only those hits).

Legal / contracts RAG Subject-scoped retrieval Cite-only answers Workers AI embeddings Vectorize Pages Functions Empty-hit fail-closed

Grounded retrieval over a contracts shelf is the core proof — the same pattern as production RAG stacks, scoped to literacy questions a small shop actually asks before signing.

  • ProblemMSA on the table; indemnity / IP / liability caps opaque
  • ConstraintCite retrieved snippets only; empty hits → stop / escalate
  • DemoPublic contract-literacy sample · edge embed → Vectorize → cites
  • ResultClause-aware hits or an honest empty set — never invented legal text

Interactive

Query the public contract-literacy sample

Public sample only — not legal advice. Zero hits means stop; don’t invent a clause.