AI Engineering · Keith Gonzales, PhD
Make your systems talk to each other.
Your answers are scattered across your CRM, call recordings, support desk, and docs, so your team digs through five tabs to answer one question. I pull them into one layer your team just asks, in plain English.
A healthcare communications platform had a year of customer data spread across its CRM, call recordings, and customer-success tools, and a team that opened four systems to answer one question about an account. As the solo builder, I connected those systems into Slack assistants the team asks in plain English, plus automated daily reports they actually open. Every number is computed in code, so the team can trust the answer. What used to take a manual dig now shows up in Slack the moment they ask.
What I build
Not dashboards nobody opens. Systems that live where your team already works and answer the questions they actually ask.
Slack bots that do real work
Bots that answer pipeline questions, run QA on other bots, flag work that's slipping through the cracks, and post the numbers your team asks for, inside the Slack you already live in.
MCP servers for Claude
Custom MCP servers that plug your internal data directly into Claude, so your team asks questions in plain English and gets answers grounded in your own systems.
System interconnectivity
Read-only pipelines that make your CRM, call recordings, customer-success data, and billing agree with each other. One source of truth instead of five open tabs.
Human in the loop, by design
Your team's knowledge makes the system smarter. My systems ask when they're unsure, remember your answers permanently, and treat your corrections as the highest tier of truth. You stay the expert; the software learns from you.
QA & eval harnesses, built into everything
This is where most AI projects fail: they demo beautifully, then quietly fall apart in production. Every system I ship comes with its own QA harness. Trust tests challenge the AI with false information to make sure it holds firm. Cross-agent verification has a second agent independently re-check the first one's claims against your source systems. And every number a human sees is computed by code; the AI only formats it. Quality assurance isn't a phase at the end. It's built into what I build, from day one.
Access on Monday. Live by Friday.
Most projects go from first API key to deployed software in under a week.
You send a problem list and read-only keys. Your eng team scopes exactly what I see. That's the whole kickoff.
I build against your live data with automated QA checking every output along the way.
Working software in your Slack, your Claude, or your browser. Iterate from real usage.
Why so fast? The scaffolding already exists. Years of production patterns, from integration clients to safety guards to QA harnesses, mean your project starts at mile twenty, not mile one.
No meetings required. Happy to meet, but I don't need to. API access and a problem list are enough. Everything I touch is read-only unless you say otherwise.
About
Twenty years of building. Now at full speed.
I'm Keith Gonzales, a PhD neuroscientist turned product leader turned AI engineer. I've spent fifteen-plus years as a product manager at some of the biggest tech companies in the world, been a Director of AI at a multibillion-dollar company, and founded two venture-backed startups of my own.
That mix is the point. I'm not only the person who builds it. Because I've been the product leader and the founder too, I can sit with your executives, your operators, your data engineers, and your ML team and speak all of their languages, then translate between them to get at what you actually need.
How your data is handled
The failure I hear about most is a tool nobody could trust with real data. Here's the default I work to.
I connect with read-only credentials your team scopes. I don't write to or change your systems unless we explicitly agree otherwise.
The software runs on infrastructure you own, or that I stand up and hand over to you. Your data is never pooled across clients or used to train anything.
I pull only what the problem needs and keep only what the tool needs to run. No quiet hoarding of your data.
Handling PHI or other sensitive data? I'll sign a BAA and work within your security and compliance requirements.
Bring me one problem.
Tell me what you're trying to solve and point me at your data. If I can't solve it fast, I'll tell you before you spend a dollar.
Prefer email? keith@hiddenfacets.com