christine.nguyen
Enterprise AI enablement, transformation, and adoption lead. I get fuzzy AI asks down to the real problem, then into working systems and capabilities teams are excited to own.
- Make it clearidentify the actual problems that need to be solved
- Make it realbuild a first version people can react to
- Make it safeput approval and uncertainty where they matter
- Make it stickleave teams ready to own and improve it
Based in Austin. Looking for Head of AI Enablement, Director of AI Transformation, or AI adoption lead roles—but don't let the titles lock you in.
Get us all loopy.
I love the moment a team stops debating AI in the abstract and starts learning from something real.
- Human in the loop. People bring judgment, taste, curation, and leadership. The interesting design question is where that judgment matters most.
- Feedback loop. Put the first version in front of the person doing the work. Their input makes the requirements sharper than another meeting would.
- Continuous improvement loop. Models change quickly and people learn just as fast. Build the system so both can keep getting better together.
Technology only matters if it fits the real work, the humans behind it, and the way they actually operate.My perspective on GenAI
I get energy from moving between the person doing the work, the leader making the call, and the engineer making it real. Eighteen years of user research, service design, and product design help me hear what each one means; building the first version gives us something concrete to translate.
Core capabilities
I move between the room where a messy AI idea is being debated and the room where a team has to make it real. I find the actual problem, shape the workflow, build enough to learn, and help the team own what comes next.
AI translation
I turn a fuzzy “we should use AI” request into a problem people can agree on, the decisions they need to make, and a shared picture of what good looks like.
Workflow design
I listen to the people doing the work, map what really happens—not what the process says happens—and redesign the touch points around them.
Prototyping
I make the first working version so people can react to something real. That gives engineers sharper requirements and empowers leaders to make better decisions.
Responsible AI and governance
I put human judgment where it changes the outcome: approvals someone owns, visible uncertainty, and a clear route for exceptions.
Enablement
I co-build with the people who will run it, teach the why as well as the how, and leave behind a capability they can improve themselves.
Strategy and leadership
I connect patterns across demand, product, data, and platform work so leaders can choose what to start, stop, sequence, or share.
The detailed tool-by-tool inventory, with an honest ship-it / direct-it / literate scale, is on the About page. The work below shows these capabilities in practice.
Selected work
AI intake
Cut a 4-month process to about 1 weekfrom the first AI request to a working prototype—roughly 94% faster, start to finish
Intake was the most manual part of my job. I turned my discovery method into an agentic process: a recorded conversation becomes written requirements, the right people sign off, and the team is testing a first version the following week. Then I made it self-service, so I scaled myself: colleagues can start an intake without me.
- My role
- Designed, built, and shipped it
- Scale
- 23 practices, multiple internal functions, and one global institute
AI portfolio strategy
Steered leaders' GenAI excitement toward the data work it needed first176 of 279 AI requests needed data and platform foundations before GenAI could help
Everyone was excited about GenAI, and the investment followed. Instead of pushing back on that excitement, I used it: sorting every request by what the AI would actually have to do showed five underlying needs—mostly data and platform work—that had to come first. Leadership used that view to change how it invested and planned its roadmaps.
- My role
- Sole author of the analysis
- Scale
- 279 requests across the firm, plus the architecture review and product-risk queues
Applied machine learning
Cut client effort for taxonomy mapping from 2.5 weeks to 2 dayswith every uncertain answer still checked by a person
86% of clients said their biggest pain point was mapping one taxonomy to another, so we built an application for it. When better prompting stopped improving the results, I led the team to train our own model: it handles the clear cases, an LLM handles the uncertain ones, and an analyst reviews what's left.
- My role
- Group product leader, discovery to handoff
- Scale
- A reusable, firm-wide component: 23 practices and their clients
AI platform strategy
Talked six teams out of a new platform and onto the one the firm already hadno engineering team needed, and the shared credential fixed first
A research institute wanted six agents and a brand-new platform to run them, with no engineering team to build it. Underneath sat one shared credential and no way to track usage. I made the case to reuse the firm's existing data platform and AI gateway instead, fixed the credential first, and got six sponsors to agree to the plan.
- My role
- Prioritization: discovery through the decision
- Scale
- Six sponsors at a research institute and a practice
Human-in-the-loop governance
Editors who can push back, with evidence in handa report's 165 facts checked in minutes, and the 45 that didn't support its argument flagged
Leadership asked for an AI drafting tool. Talking with the editors showed bigger pain points upstream: they lacked the evidence to push back on authors. So the first thing we shipped checks an outline against its evidence—solving the immediate pain and laying the groundwork for the drafting strategy.
- My role
- Product lead; built the plugin's skills, co-developed with an engineer
- Scale
- Editors, writers, and analysts at a research institute
AI operating model
A safe way for teams to train their own AI modelsa small trained model first, the expensive LLM only when it's unsure, and a person on the hard cases
I set up a safe path for teams that had reached the limit of prompting: the trained model handles clear cases, the LLM handles uncertain ones, and a person reviews the rest. By design, that's roughly three quarters fewer LLM calls.
- My role
- Product lead; wrote the operating model
- Scale
- One approved path to train a model, where there had been none
Agent orchestration
Turned the task everyone dreads—building a portfolio—into a one-week projectagents did the digging; I made every call
Nobody likes building a portfolio, me included. I wanted to know where I fit in a fast-changing AI job market, so I had agents recover thirteen months of my work from six systems, then checked every claim myself. I rediscovered work I'd forgotten, and the story got sharper.
- My role
- Operator and reviewer
- Scale
- Thirteen months of work across six systems
Open source
Most of my work is confidential, so I rebuilt the methods behind it from scratch—made-up data, no proprietary code—and I'm open-sourcing them so others can use them.
Intake method
The intake methodfrom a recorded conversation to technical requirements in one pass—enough to build a proof of concept within about a week
This is what gets product managers excited: far less back-and-forth to reach alignment, and something real for users to react to within days. You can inspect the workflow, the approvals, and the weak spots the evaluation exposed.
- My role
- Designed and wrote it, alone
- Data
- Synthetic only; MIT licensed
Portfolio method
How I built this portfoliothe evidence method behind this site, rebuilt with made-up data so anyone can run it
Most people write a résumé from memory and go looking for proof later. This flips it: write down what you want to say, let it check each line against your own records, and nothing goes public until you say yes. It even tells you where it got one wrong.
- My role
- Designed and wrote it, alone
- Data
- Synthetic only; MIT licensed
My POV
Publish confidence thresholds, not accuracy headlines
"It's 95% accurate" is unfalsifiable in the way that matters. Publish the curve and a threshold per workflow instead, and say what happens to the other five percent.
Three gates: human approval belongs inside the workflow, not in the policy deck
The gate took an afternoon to build. Deciding what to route through it took months, and I got it wrong first.
How to read your own AI portfolio
Sort requests by what the model does, not by what the requester called it. Then look for the bucket your taxonomy can't hold. That's where the money is.
Intake as a system
"How do we prioritize?" is the wrong first question. The first question is how do we see. You can't rank demand you haven't captured in a form that can be compared.
Let's get loopy
Working software is now how a team communicates, decides, and validates. That turned my assembly line into three loops — and a person belongs in each one.
Oversight has a capacity
A field report on making "human in the loop" real: where gates belong, why escalating everything makes systems less safe, and how to publish a risk–coverage curve.
Steps to get loopy
A working first version makes the room more curious—and the conversation much better. I use five short loops to listen, build or reuse, and test with the people doing the work. The part I care about most is what each loop teaches us: the next best set of questions to ask.
What could your team learn this week if the first version were something people could try?
Where
Christine Nguyen · Austin, Texas.
Open to Head of AI Enablement, Director of AI Transformation, and AI adoption leadership. Don't let the titles lock you in: if the work matches my skills, I'm open to it.
If you've read this far, we should probably talk.
Bring me the AI problem your team is still trying to explain.
christineqnguyen@gmail.comClick to copy.