AI Architect · Texas Triangle (San Antonio / Austin / Dallas–Fort Worth) · Website · Email · LinkedIn · X

I design agentic systems where a model can propose and cannot grant itself permission. Financial services trained the consequence. Founding trained the ownership. The public repos are the parts a reviewer can inspect.

Current: Founder, Apex AI|ML. Prior: Minotaur Consulting · Office of the CIO

Three practical products. Three different business problems. One architectural rule: the model can propose; the system determines what is authorized.

The operating pattern: understand the workflow → bound the intelligence → require human authorization → verify current evidence → preserve the outcome. These are reference implementations, not claims of paying customers or enterprise production deployment.

Engineering explorations targeted at enterprise AI architecture and forward-deployed delivery; none are claimed as completed portfolio evidence.

• Azure OpenAI / Azure AI model workflows: multi-turn history, system/developer instructions, image understanding and OCR; compare model families and enforce data-handling, cost, and evaluation boundaries.
• Production SaaS architecture: Next.js on Vercel, AWS App Runner, and Clerk authentication/user management; subscriptions, tenancy, and operational readiness would need independent implementation and security validation.
• Cross-cloud cybersecurity analyst: prototype an MCP-connected analyst workflow on Azure and GCP with approval-gated actions and auditable evidence.
• Multi-agent orchestration: compare coordination patterns, failure modes, tool permissions, context budgets, and human escalation before choosing a framework.
• Claude Code / agentic development lifecycle: evaluate AI-assisted development with tests, source control, human review, and accountable delivery; also compare Codex and other coding harnesses.
• Linux system hardening: reproducible baseline, least privilege, service lockdown, audit checks, rollback, and validation on a disposable test environment.

What I bring to client-facing AI delivery: verifiable shipped code; Python, cloud architecture and MCP integration; bounded tool authority and model evaluation; founder-level ownership; and translation from an executive business problem to a reviewed technical decision. No cloud production claims are implied for the planned items above.

These eight projects show depth in governed AI and breadth across capital markets, energy, supply chain, and cybersecurity assurance. BALLAST makes operational trade-offs visible; AEGIS Evidence makes governance claims and unresolved control gaps inspectable.

Monster Light and Monster Desk remain supporting paper-boundary examples alongside Monster Heavy.

A recommendation is not permission. Approval is bound to the exact action. Current evidence is checked again at execution. If the control cannot be shown, the claim is not made.

Business risk crosses departments. The workflow should connect the pieces.

I use Gradio to turn Python workflows into interactive interfaces that business reviewers can explore. Gradio's team reported more than one million monthly developers in April 2025, and its integration with Hugging Face Spaces supports rapid demonstration and sharing.

Three lines create the interface. The engineering behind it makes the workflow worth using.

import gradio as gr
demo = gr.Interface(fn=analyze_risk, inputs="text", outputs="text")
demo.launch()

Illustrative interface example: analyze_risk is an existing Python function. BALLAST and AEGIS Analyst use richer layouts and explicit workflow controls.

• BALLAST — Gradio supply-chain control tower: energy shipping routes and Taiwan semiconductor dependencies make concentration, single-source exposure, and disruption visible. Operators compare baseline and what-if scenarios, inspect risk drivers, and explicitly approve disruptive mitigation simulations.
• AEGIS Analyst — Gradio cybersecurity console: seeded security alerts become analyst briefs, ATT&CK mappings, and containment proposals. Host isolation requires explicit approval before the sandbox returns a simulated result. Its AEGIS Evidence companion uses Streamlit to expose control declarations, human acceptance, unresolved gaps, and verifiable evidence archives.

Together, these workflows connect supplier resilience, energy exposure, semiconductor dependencies, security operations, and accountable decisions. The demonstrations use seeded or synthetic scenarios; mitigation and containment are simulated. The business value is making the risks, trade-offs, and authority boundaries inspectable before operational changes.

No hosted production systems are implied. Install each repository's documented dependencies in its own virtual environment, then use the relevant local entry point.

Streamlit applications (from a repository containing app.py, such as the AI Governance Helper or compatible dashboard):

python -m streamlit run app.py

If your Python environment is activated and streamlit is on PATH, the shorter command also works:

streamlit run app.py

Gradio applications: follow the relevant repository's documented startup command; BALLAST and AEGIS Analyst have their own entry points and workflow settings. See the READMEs for fixtures, optional API requirements, and safety limitations.

Three interactive local applications combine Massive historical daily market evidence, Python-calculated metrics, and bounded CrewAI research. Structured claims are checked against the supplied evidence; interpretation still requires human review. Agents have no order tools. Only the paper floor records simulated fills, after explicit local confirmation and account checks.

• Capital Markets Research Desk pairs a research brief with a fixed-rule chronological holdout, benchmark comparison, and visible cost assumptions.
• Investment Gems applies transparent momentum, volatility, and liquidity hurdles to a selected equity/ETF universe.
• Paper Trading Floor retains simulated cash, positions, and an evidence-linked SQLite fill ledger across restarts.

Each repository includes screenshots, setup instructions, tests, and explicit limitations. Data are end-of-day historical observations; price returns exclude dividends. Energy equities and ETFs are proxies, not ERCOT power or Henry Hub spot feeds. There is no broker connection or real-money execution path. The original static fixture views remain documented as legacy examples.

Evidence scope and confidentiality

Confidential client material is not published. Public FinOps repositories contain modeled scenarios with documented formulas and assumptions.