Hanze Li is one of Cresta's founding Forward Deployed Engineers. With a background that spans hardware, machine learning research, software engineering, and student government, he embodies the "connecting the dots" philosophy that defines the FDE role. We sat down with Hanze at the first-ever FDE team offsite in Salt Lake City to hear his story.
Tell us about your background — it sounds like quite a journey.
It really is a journey. I started coding when I was eight years old in China — my parents sent me to an after-school coding club, and I hated it. I told myself I'd never touch code again.
Fast forward to university in Canada: I enrolled in a program I thought wouldn't require much programming, but a year or two in I completely fell in love with coding. From there I went in many directions — I did an FPGA and Verilog hardware internship, sitting in a lab testing real silicon chips. I said "no, not for me," and went back to software. Then I got drawn into ML research, published three or four papers, thought maybe I'd do a PhD in foundation model training. Halfway through my undergrad thesis I said "no, that's not me either."
I ended up doing a master's in AI at the University of Toronto, and in parallel I co-founded with many friends what is still one of the largest machine learning student clubs there: University of Toronto Machine Intelligence Student Team (UTMIST). I also ran campaigns for the governing council as a student representative. None of it seemed to connect at the time. And then I joined Cresta as an FDE, and suddenly every single piece of that background was useful.
How do all those experiences connect to the FDE role?
There's a Steve Jobs analogy I keep coming back to — his famous "connecting the dots" theory. He worked on tech, hardware, studied calligraphy, and traveled to India. It all looked like wasted time, until it didn't. For FDE, it's the same: you don't need to be a senior staff backend engineer who knows every inch of infrastructure, and you don't need to be a research scientist developing foundation models. But if you're above average at all of those things combined — coding, machine learning, customer communication, even public speaking — you're really, really good at FDE.
Every FDE is almost like a CTO in a startup company: strong on tech, able to handle customer engagement directly, but preferring to work on the technical side rather than purely the business side. You need to build trust with the customer and work out a good plan together to make things happen. It's not just blindly following whatever the customer wants — it's a conversation, based on trust, experience, and skill.
The common trap is the traditional software engineering mindset: a PM gives you a ticket, you say "I'm on it, two weeks, high-quality delivery." As an FDE, that might not be the best approach. You have to understand what the customer really wants, come up with options, and drive the prioritization. Option one: change it properly, two weeks, perfect. Option two: minor change, eighty percent of the benefit, minimal effort. Option three: do nothing — it's still good enough. Most customers choose the middle option. As an FDE, you're the one who creates that menu.
What does a typical day look like?
The official answer is 1/3 AI agent development, 1/3 customer engagement, 1/3 product-related work. In practice it shifts over time.
When I first joined, probably more than half my time was on AI agent development — getting up to speed on our infrastructure, tools like Claude Code, all the agentic engineering primitives. Those are efficiency multipliers that save you enormous time later.
After about six months, my split has flipped. Less than a third is coding now. I can build a demo-ready version of even complex use cases in a day or two. The rest is testing, evaluation, and customer engagement. Even though I'm not an FDPM, I'm on the business side too — I'm the representative explaining the technical world. When a customer gives me a request I can say, "This feature seems complex but is actually easy," or "This seems simple but has significant backend implications." That back-and-forth, building real trust, is a huge part of the work.
One thing that really inspired me after joining Cresta: a lot of problems that sound technical turn out to be solved by interpersonal skills. A customer reports a bug. It's not really a bug — it's a feature we haven't delivered, or something we can work around. The skill is knowing how to present the options clearly, let the customer make an informed decision, and move fast on the right thing. In the LLM world there's always some hallucination no matter how carefully you handle edge cases, so you really need to prioritize rather than just fix.
You work with Brinks Home, one of Cresta's customers. What does that relationship look like?
Brinks Home is a very special client. I've worked with quite a few customers, and what makes Brinks unique is the level of trust — we move together. It's not "Cresta builds something, then Brinks decides whether to buy it." It's more: "We have something new. We're confident it works, but no one has adopted it yet. Do you want to be the first?" And Brinks always says yes.
The metaphor I use is omakase-style. In a traditional omakase restaurant, you trust the chef completely; they bring you the best they can offer, and your job is to try it with an open mind. Brinks extends that same trust to us. We make sure everything we bring is safe, tested, and the best we can give. And because we've done well in the past, they trust us for the next step.
One recent product we just launched with Brinks is called Agent Operations Center — it's a human-in-the-loop system for AI agents. Instead of replacing human agents, it gives AI agents simple, repetitive work and lets them raise their hand when something gets complex, likea discount request, an unusual edge case, anything you don't want AI deciding alone. The AI agent surfaces the context, a supervisor in the backend says "give twenty percent off" or "push back," and the agent continues the conversation seamlessly. This is very similar to real human agents where they will need a senior supervisor to help from time to time for complex or sensitive matters.
As the lead FDE on that launch, you have to think through every edge case. What if there's no supervisor online? What if the one who got paged forgot to reply? What if the AI agent misunderstands the instruction? The first time you do something, all of those questions are live questions. The second time is much easier because you have experience and a mature product. But the first time, you have to think through everything, and that's one of the most exhilarating parts of the job.
How do FDEs approach testing and evaluation?
This is something I feel strongly about. Building an AI agent is a research problem — you can always get something quick-and-dirty working, but you never know if it's the best version for your client or vertical.
For example: always showing empathy might be good in healthcare, but in fintech, if someone owes two or four thousand dollars to a bank and the AI agent is being overly empathetic, people have mixed feelings. Different verticals, different customers want different things. A certain function tool or prompt that works perfectly in one vertical might fail completely in another.
The answer is data. Build a dataset, build a benchmark, validate that your choice of function tools, prompts, and reusable agent blocks is actually the best choice — and that takes time. It's not "I tried this, it worked for this customer, so it must be best." You need to run proper evaluations, compare options A, B, and C, and use a data-driven approach to see which is genuinely better.
We're now using GPT-4.1 for some workflows. When we upgrade to a newer model version, are our prompts still best practice? We don't know — unless we have a benchmark. The goal is: for this vertical, with these prompts, with this base model, the conversation still performs well. Then we can quickly iterate, add new model capabilities, and trust that regressions still pass.
That's part of Cresta's secret sauce: we have so many customer engagements and data points that only we can build this kind of benchmark for enterprise AI agents. Vibe coding works perfectly for demos and toy agents, but in production, you can never trust it without enough data and evaluation behind it.
What's your advice for someone considering an FDE career?
Try different things. Hardware, software, machine learning research, student government, entrepreneurial work — everything can connect. You never know when an unusual experience will help you build trust with a technical leader, or help you solve a problem in a more systematic way.
The mindset shift is important too: be the owner not the executor. Don't think "PM gave me a task, I'm on it." Always understand the big picture, come up with different options, and use a data-driven approach. The code itself is just a way to make things happen. The architectural design, the business model, and the relationship with customers may matter even more now — and even more in the future.
I've interviewed over thirty people for FDE roles. Many say "I built a chatbot at home." But when we dig in, they've missed a huge portion of what a real production chatbot needs. You can always read blogs, but without real production experience across different use cases and verticals, you don't get the full picture. There's a lot of hidden knowledge in any industry — things that seem like common sense only if you've been inside. You don't know what you don't know.
Where do you see enterprise AI heading in the next few years?
In two or three years, I hope nobody has to wait on hold for an hour to talk to an airline. You share your intent, the system remembers your account, and things get resolved in a few minutes. Then we can spend more time with our families, with our books, go skiing — and focus on what really matters.
That's what FDE work is building toward. Not just clever demos, but AI agents that are reliable, trustworthy, and genuinely better for the people using them. I feel none of my background got wasted when I joined Cresta. All those dots connected in a way I couldn't have planned.
Curious about Forward Deployed Engineering or building enterprise AI agents? Cresta is hiring FDEs across different domains. Check out our careers page to learn more.
Watch Hanze's full interview here.
