Deion Wells Ross

AI Solutions — from problem to adoption

I turn operational problems into AI-enabled workflows that change how work gets done.

I work at the intersection of business and technology — identifying where work breaks down, designing AI-enabled workflows to fix it, integrating the technical solution, and helping people actually adopt it. Not an introduction to AI. A change in how the work happens.

View selected projects Get in touch

Selected work

Selected Projects

A few representative engagements — the problem, what was built, and the outcome for the business.

01

AURI Finance

  • Agentic AI
  • Multi-Agent Systems
  • Human-in-the-Loop
  • Claude API
Problem
Finance teams rely on manual workflows, deterministic accounting systems, human judgment, and financial controls to run processes like month-end close — simply adding an LLM doesn't make that trustworthy. The real question: how can specialized AI agents do real financial work while deterministic systems keep the math correct, independent controls catch agent mistakes, and humans keep authority over consequential decisions?
Solution
I built AURI Finance, an agentic finance department for a fictional B2B SaaS company. Seven Claude-powered agents — bookkeeping, accounts payable, accounts receivable, payroll, controller, FP&A, and CFO — handle the judgment calls, using scoped tools to investigate exceptions and escalate uncertainty instead of guessing. Two deterministic (non-AI) components, an orchestrator and a dashboard publisher, calculate every number and sequence the workflow. A close moves through the specialist agents, an independent control re-check, FP&A/CFO analysis, and a human approval gate before the period closes — so no agent's self-report is ever taken at face value.
My contribution
I defined the problem and the bar this system had to clear — that agents do real financial work rather than just demonstrate intelligence — and directed the architecture: specialized agents, deterministic services, scoped permissions, independent controls, and human approval gates. I used Claude and Claude Code as my agentic development environment to build, test, debug, and document the system end to end; this project is itself a demonstration of AI-native development.
Business outcome
Tested end to end against a fresh synthetic database the agents had never seen: 6 of 6 seeded anomalies found, 55/55 tests passing, $0.00 trial balance discrepancy, a 202-row audit trail, and a required human approval before close — for about $7.10 in total Claude API spend. Real agent mistakes surfaced during live testing; the independent controls caught them before they could complete a close.

02

AURI Access

  • Google Drive
  • Apps Script
  • Claude API
  • Permissioned Access
  • Audit Trail
Problem
Companies increasingly want AI agents reasoning over their internal knowledge — but most organizational knowledge is scattered across departments with no consistent access controls, and once an AI agent is reading it, there's usually no record of what it was shown versus what it was correctly withheld. For a regulated business, "the AI read our files" isn't good enough; leadership needs to know exactly what was accessible, to whom, and why.
Solution
I built AURI Access, a permissioned intelligence layer on top of a live Google Drive environment for a fictional company, Meridian Analytics, spanning Sales, Marketing, Finance, HR, Operations, and a restricted Leadership tier. Drive itself acts as the permissioned source of truth; a deterministic Apps Script layer handles all retrieval and enforces access boundaries in code, never in a prompt; Claude synthesizes answers only from what that layer is permitted to return. Two audit mechanisms run in parallel — Google's native Drive activity log for compliance-level access tracking, and a custom semantic audit log that records what was asked, what was returned, and what was deliberately withheld, which is the more meaningful artifact for a regulated audience.
My contribution
I designed the permission architecture, built out the seed Drive environment including deliberately planted knowledge-hygiene problems (a stale sales playbook, duplicate onboarding guides, an orphaned vendor SOP with no clear owner) to stress-test retrieval against a realistic, messy organization, and built both audit layers. I verified live document retrieval and Claude synthesis end-to-end before preparing the system for deployment.
Business outcome
A working demonstration of how a company can let an AI agent reason over sensitive internal knowledge without granting it blanket access — access stays scoped and provable, not assumed. The semantic audit log in particular gives a compliance team something most AI deployments can't: a record of what the system chose not to show, not just what it did.

03

Placeholder Project Three

  • Opportunity assessment
  • Process mapping
  • Roadmapping
Problem
Leadership wanted to adopt AI but had no framework for deciding where it would actually create value.
Solution
Ran a structured opportunity assessment, scored candidate workflows on impact and feasibility, and delivered a sequenced roadmap.
My contribution
Facilitated stakeholder interviews, built the scoring model, and produced the prioritized implementation plan.
Business outcome
Placeholder outcome — e.g. a prioritized portfolio of initiatives with a clear first project underway within the quarter.

Background

About

My work spans AI implementation leadership, forward-deployed engineering, management consulting, workflow automation, and customer adoption. The through-line: I sit between the business problem and the technical solution, and I stay involved until the solution is actually part of how the team works.

In practice that means running discovery to find where work breaks down, designing AI-enabled workflows around those problems, building or integrating the technical pieces, and coaching people through the change. I care as much about adoption as about the build — a solution nobody uses solved nothing.

I hold an MBA from Nebraska Wesleyan University, which grounds the technical work in how organizations actually make decisions, allocate resources, and measure value.

Method

How I Built This

This site is itself an example of the work. It was built by directing an AI coding agent through a defined process — the same way I approach client solutions.

  1. 01

    Defined the audience and objective

    Named the visitors this site is for — hiring leaders, recruiters, potential clients, collaborators — and the single job it needs to do: show that I can translate a business problem into an AI-enabled solution.

  2. 02

    Translated the brief into architecture

    Wrote a structured brief — purpose, audience, scope, tone, constraints — and turned it into an information architecture and a phased build plan before any code was written.

  3. 03

    Directed iterative AI-assisted development

    Built the site by directing Claude Code as a coding agent: specifying each section, reviewing the output, correcting course, and iterating in passes rather than one prompt.

  4. 04

    Reviewed and refined the code

    Read the generated code, tightened structure and semantics, kept the dependency footprint minimal, and made deliberate calls on what to keep, cut, or rework.

  5. 05

    Tested and deployed

    Checked responsiveness, links, accessibility, and performance, then deployed on Vercel with continuous deployment from the main branch.

Contact

Contact

If you're working on something where AI could change how the work gets done, I'd like to hear about it.