For companies starting AI from zero, or stuck at a stalled pilot
I founded and ran the AI program at a 400-person company: 15+ tools in production, governance included. Now I do the same for clients: work already shipped, easier to start and easier to stop than a full-time hire.
Every engagement starts with the Diagnostic.
The gap
The demo works. Then it meets your identity provider, your security review, your data retention policy, and the person who has to own it next quarter. That is where most AI programs quietly stop. If you have not started yet, it is the part worth planning for first.
We built something great in a hackathon. Six months later it is still not live.
Fit
Good fit
Not a fit
Services
Most engagements start with one and grow into the next as the first thing proves out.
Zero to an AI program that sticks.
Your first AI strategy, an LLM data-handling policy, and an approved-tools framework, plus the governance to hold them in place. This is the work that keeps AI adoption from turning into shadow IT you find out about later.
Tools your team opens and keeps using.
MCP servers and LLM applications wired into the systems you already run: enterprise SSO, human-in-the-loop approval gates on every write action, full audit logging. Built to hold up in front of your security team, not just in a demo.
Most stalled AI projects are actually stalled data projects.
Warehouse design, ingestion, and modeling in Snowflake, dbt, and Airflow, with CI/CD, monitoring, and alerting from day one rather than bolted on later.
Your non-engineers can build. Safely.
A guardrail framework and review process that lets analysts, marketers, and operators ship internal tools without shipping risk alongside them. Includes hands-on hackathon facilitation and technical coaching.
Approach
Every engagement starts with the Diagnostic below: a fixed-scope look at where AI pays off in your systems and what is blocking it. You keep the roadmap whether or not we work together after.
One tool in production in weeks, not a deck in months. Real users, real data, real approval gates, real audit trail. Momentum comes from a thing people use, not a thing people approve.
Documentation, patterns, and enablement so your team owns what was built. The goal is for the engagement to end cleanly, not for you to need me indefinitely.
I get inside your systems and tell you, in writing, whether AI is worth building here at all.
Timeline and fee are quoted after a short scoping conversation.
Start the DiagnosticStarting an AI program, or unsticking one?
Start a conversationSelected work
The examples happen to be consumer products, marketing, and direct sales; the underlying work is the same in any industry: auth, approval gates, data, and audit trails. Client names are withheld; a reference call is available on request.
Legal required that every word users see be pre-approved. So the LLM classifies questions but never generates answers: it matches natural language against an approved FAQ database and returns the exact approved text, in English or Spanish, with confidence scoring and a human fallback. In production on Aurora MySQL with IAM auth.
Authored the company's first AI strategy, LLM data-handling policy, and approved-tools framework. All three now inform company-wide governance. Also serves as architecture reviewer for the company's AWS Bedrock AgentCore platform: auth model, rate limiting, secrets management.
Production MCP servers with Entra ID OAuth and human-in-the-loop approval gating on every write action: ticket creation, ad budget changes, user group management. Logging, auditing, and retention owned in-house, so LLM usage data never leaves company infrastructure.
Google, Meta, TikTok, Microsoft, Pinterest, and GA4 behind one interface, with approval-gated budget and status controls so the marketing team can move spend from a conversation instead of six dashboards.
Custom ingestion, Airflow, dbt, and Snowflake, with CI/CD, monitoring, and alerting. In a prior role: redesigned an ELT pipeline for an estimated 50% Snowflake cost reduction and consolidated 3,000 analytics models down to 200 with consistent cross-department definitions.
A 250-key LLM creative-tagging dictionary that turned unstructured ad creative into modelable features, making image performance predictable at roughly 96% accuracy. Plus an AI code-review bot, live on every merge request, that pre-screens changes before a human sees them.
Built in house
The practice runs its own product. firstPass connects ad platforms to Claude, ChatGPT, Slack, and Teams so marketing teams can ask questions of live campaign data instead of exporting it. Read-only access, no raw ad data stored. It is also the reference implementation for how these engagements get built.
Principles
Every system I build assumes it will eventually be audited, inherited by someone else, and asked to explain itself. That is not overhead; it is what separates a tool that ships from a tool that gets quietly turned off.
About
I lead firstPass. I founded the AI program at a 400-person consumer-products company: 15+ LLM-powered tools shipped, org-wide enablement and governance, and architecture direction for an AWS Bedrock agentic platform, all on a deep data-engineering foundation. Before that, I was an analytics engineer building warehouses and pipelines at scale. B.S. Information Systems, BYU Marriott School of Business.
Agents & AI
Data & platform
Certifications
Why this model
A senior AI or technical hire is a six-figure bet before you know if it pays off: months to find the right person, more months to onboard them, and results only after that. I am not asking you to make that bet. The work on this page is already in production; you are hiring what you can see, not what an interview suggests.
I also start in days rather than a hiring cycle. Every engagement opens with the Diagnostic, and the roadmap it produces is yours to keep whether or not we continue. If it does not make sense to keep going, you walk away with something useful, not a sunk cost.
And when an engagement ends, it ends. I document what I build, hand over the patterns, and stop billing. If you later decide the right answer is a full-time hire, what I leave behind is designed to make that person’s first month easier, not something they have to reverse-engineer.
If you are weighing this against a name-brand firm instead of a hire, the trade is simple: a bigger logo, or direct access to the person doing the work, at a price that reflects one person’s time rather than a staffed team’s.
Questions
You can, and for some situations that is the right call. If the engagement proves the role deserves a full-time person, that is a good outcome, and nothing here forecloses it. I would rather you make that call with a working system in front of you than a job posting.
For the engagements I take on, yes. They are scoped so one senior person can carry them end to end, which is also why nothing gets lost between a strategist and an implementer. When a project genuinely needs a team, I say so in the Diagnostic and help you scope what to hire instead.
Everything I build lives in your repos, on your infrastructure, documented as I go: nothing exists only in my head. If something took me out tomorrow, you would still own the code, the docs, and the roadmap, and a competent engineer could pick them up. That is not an accident; it is the same discipline that makes the handover at the end of an engagement work.
Engagements are scoped to a problem, not to headcount. The program I built covered a 400-person company end to end, and current client work is for a considerably larger global organization. What changes at a bigger company is the handover: the keys go to a team instead of a person.
Tell me what you are building, or what is stuck. If it is not something I can help with, I will say so and point you somewhere better.
I read every message myself, usually within one business day.