About · Agentic AI engineer

Joshua L. Coleman

I build agent tooling for work where a wrong action is expensive. I lead operations for two mission-critical AWS data centers, build the MCP servers and guarded agent workflows used in that work, and teach engineers and technicians to run them.

Get in touch LinkedIn GitHub

Operator and builder

I studied electrical engineering and built my career in critical infrastructure: battery validation at Romeo Power, then data center engineering operations at STACK Infrastructure and AWS. Today I'm a Lead Data Center Engineering Operations Engineer at AWS in Northern Virginia, running two live facilities that carry hyperscale and machine-learning workloads. That means preventive maintenance, controlled procedures, change management, incident escalation, and troubleshooting electrical and control systems.

That job is why I build agents the way I do. In a data center, a procedure that skips a step can drop a load, so the tools I hand a model are narrow on purpose: scoped operations instead of open file writes, per-tool and per-path permissions, allowlists, backups and undo, and a human approval before anything leaves the building. I use AI coding agents to move fast, and I own the architecture, the code review, the integration and the validation.

Most of this is forward-deployed work in practice. I sit with the people doing the job, find the step that hurts, ship a tool that fits their process, and train them until they run it without me.

Teaching operators to run agents

I hold classes for engineers and technicians on running agent workflows in their day-to-day operations. I built the course myself: 13 slides on Kiro and MCP with applied labs, covering procedure editing, reporting, and the guardrails that keep an agent inside its lane.

The audience is people who maintain switchgear and generators, not software engineers. The class is built so they can use the tooling on real maintenance work, and so they know where it should stop and ask.

What I bring to a deployment

Tool design with limits. MCP servers that expose the operation a user needs and nothing wider.
Human checkpoints. Approval gates on anything external or hard to undo.
Failure handling. Explicit failure states, rate-limit handling, reconciliation, and tests against mocked APIs.
Adoption. Training material and live classes, so the tool keeps being used after I leave the room.

Selected work

Agent tooling, full-stack products, and forecasting systems.

Python · MCP · Kiro

Enterprise MCP tools and guarded agent workflows

Three MCP servers in active use. mop-word makes structured edits to controlled MOP and SOP documents (step and table operations, numbering checks, backups, undo, finalization) without giving the model unrestricted file writes.

Others connect Outlook, Slack, Asana and internal knowledge workflows through narrowly scoped tools, with per-tool and per-path runtime permissions, a server-side email allowlist, and approval gates on external communications.

FastAPI · React · TypeScript · SQLAlchemy

Federal Contract Intelligence Platform

Ingests and scores SAM.gov opportunities: about 1,500 notices in the initial backfill, with profile-based ranking, Claude-assisted solicitation analysis, watchlists and pipeline tracking.

Backfill, incremental, reconcile and refresh sync modes; database migrations; UTC-safe timestamps; rate-limit handling; explicit failure states; mocked API tests with pytest and respx.

Node.js · Python · PyTorch · Chronos

ChartForge quantitative AI platform

A local-first prototype that charts 25+ years of 5-minute NQ/NDX history with a custom Canvas renderer, authenticated WebSockets, persistent Python workers and a streaming Anthropic assistant.

Time-series forecasts, regime detection, calibrated forecast intervals and no-lookahead backtests aligned to the rendered chart, with atomic reconciliation to guard against partial ingestion.

Python · MCP · FRED API

CPI Nowcast MCP server

Headline and core inflation nowcasts from 14 weighted components, market proxies, gasoline and lagged shelter data. Forecast breakdowns and Cleveland Fed comparisons are exposed as agent-callable tools and checked against published releases.

AWS EC2 · Python · Broker APIs · Stripe

Trading SaaS and automation

Subscription trading bots with automated EC2 provisioning, broker integrations, payment workflows, ensemble forecasting models, CI/CD and Monte Carlo strategy validation.

PowerShell · WPF · MCP

Operations tooling

PrintDoctor: layered printer diagnostics across port, queue, driver and device. Digital WAF: a local work-authorization tracker with a staged approval lifecycle. sticky-notes: an MCP server for persistent agent notes.

Hiring for an agentic, applied AI or forward-deployed role? Send me a note. My full work history is on LinkedIn.

History of this site

The trail, applied to itself.

2023The Prompt Files is conceived as a research site on LLMs and GPTs. First pages: a reading list of three papers (the Transformer, two vision-language pretraining papers), a prompt library of planned articles, a research page listing the Adaptive Language Weighting program, and a "Perfect Prompt Constructor" backed by a local Python service.
Apr 2024The site is committed to GitHub as a static HTML site with a mobile navigation and the photo you see above. Most library and research links are still placeholders.
2024–2025The field moves faster than the site: reasoning models, million-token contexts, open frontier models, the Model Context Protocol, and the alignment results that changed how reasoning traces are read.
Nov 2026Rebuild. The paper list becomes a continuous, filterable timeline from 2017 to the present; the library gains real guides and templates; the constructor runs in the browser; the site is served through Cloudflare. Every placeholder from 2023 is either filled in or removed.