The Tech Trek

Getting a voice agent to talk is one challenge. Knowing whether it can handle real conversations is another.

Brooke Hopkins, founder of Coval, joins The Tech Trek to discuss how teams test, evaluate, and scale voice AI agents. Drawing on her experience leading evaluation infrastructure at Waymo, Brooke explains why testing voice agents shares surprising similarities with testing autonomous vehicles.

The conversation explores why older voice assistants struggled, how better reasoning models changed what's possible, and what it takes to earn users' trust.

Brooke also shares how her engineering team works with multiple AI agents, why technical interviews need to assess AI skills, and how she balances automation with human judgment as a founder.

Key Takeaways

- Voice AI evaluations must account for interruptions, background noise, and complex conversations.
- Simulations identify failures before deployment, while production QA reveals what testing missed.
- Engineers need new ways to divide work when running several AI agents in parallel.
- Hiring should assess problem solving, system design, and how candidates use AI.

Episode Highlights

00:32 How Coval tests voice agents before and after deployment
01:26 What autonomous vehicle simulation teaches us about voice AI
03:24 Why better speech recognition wasn't enough
08:00 Could voice AI change how we use screens?
14:43 How AI is changing engineering interviews
16:43 Where a founder uses AI and where human judgment matters

A Moment Worth Pulling Out

"We had the eyes and the mouth, but we didn't have the brain."

Follow The Tech Trek for more conversations about engineering, AI, and building technology companies.

What happens when AI changes more than coding speed?

Chris Merrick, Co Founder and CTO at Omni, explains how his engineering team is adapting around AI, autonomy, product ownership, and a lighter operating model.

Omni has about 25 engineers working in groups as small as one to three people. Engineers are expected to make product decisions, move without waiting for detailed specifications, and show unfinished work during a weekly company demo day.

Those demos are also posted publicly. Customers watch them, join betas, and see features often move into production soon after.

Chris also shares how Claude Code changed development across the team. Commits to main increased roughly two to three times without comparable team growth, but the added output created another constraint: code review.

Takeaways

• Small teams can operate with less process when engineers have real ownership.
• Public demos create feedback loops between engineers, customers, and the rest of the company.
• AI can increase engineering output while moving the bottleneck somewhere else.
• Token costs still matter when AI usage scales across products and engineering teams.

Key Moments

01:15 Building AI into the engineering workflow
03:38 Why Omni works in teams of one to three engineers
06:48 The weekly demo day that replaces heavier process
10:29 Why Omni publishes unfinished product demos publicly
17:14 Measuring AI engineering productivity through commits to main
19:52 When token economics start affecting engineering decisions

One Line That Stuck

“I’m not married to the process. I’m married to the outcomes.”

Follow The Tech Trek for more conversations about how modern engineering and product teams are changing.

AI coding tools are helping engineers write code faster. But faster code does not automatically mean better business outcomes.
Vitaly Gordon, cofounder and CEO of Faros, joins The Tech Trek to talk about the gap between AI adoption and measurable results. The conversation looks at why engineering teams are spending more on AI, how enterprises are thinking about ROI, and why existing processes can become the bottleneck even when code generation speeds up.
Vitaly also explains why AI adoption inside large organizations looks different from simply giving engineers access to new tools. Teams still have to deal with security, legacy processes, risk, integration, and organizational change.
Key takeaways
• More AI usage does not automatically create more customer value.
• Engineering leaders need to connect AI spend to actual outcomes.
• Enterprise adoption requires process changes, not just new tools.
• AI ready engineers increasingly need both domain expertise and AI skills.

Key moments
00:36 Stop token maxing and start outcome maxing
02:51 Why FOMO accelerated AI adoption
06:32 Measuring ROI on engineering AI spend
09:19 Change management inside large engineering organizations
14:23 Why AI ready engineers are harder to hire
17:13 Will AI reduce engineering jobs?

Best Line
“Stop token maxing and start outcome maxing.”

Follow The Tech Trek for more conversations with technical leaders building and operating modern teams.

Technology teams are often asked to solve something before everyone agrees on what the actual problem is.

Chad Carrington, CIO at Golden1 Credit Union, joins The Tech Trek to talk about listening to business partners, defining the outcome first, and resisting the instinct to jump immediately to a technology solution.

The conversation also looks at how AI is changing this skill. As tools make building easier and faster, people still need to articulate what they actually want, ask better questions, and understand what success looks like.

What You Will Hear

• Why listening and hearing are not the same thing
• When solving 85 percent of a problem may be enough
• Why technology teams need more time with business users
• How AI is increasing the value of asking better questions

Key Moments

01:20 Listening versus actually hearing the business problem
03:47 When immediate customer impact matters more than technical debt
06:31 Bringing the business into decisions about speed and tradeoffs
09:47 Why technologists need to spend more time with business teams
13:20 How easier tools change the importance of defining the problem
18:33 Why AI is still a tool, not a substitute for expertise

One Line That Stuck

“There’s a difference between listening and hearing.”

Follow The Tech Trek for more conversations with the people building and leading modern technology teams.

AI has learned from the digital world. Physical AI brings real world data into the picture.

Doron Hazan, Director of Products and AI at Wiliot, joins The Tech Trek to explain how physical AI connects AI systems with objects, environments, and supply chains.

The challenge is not simply processing data. It is collecting accurate, current information from the physical world.

Doron explains how ambient IoT, sensors, statistical inference, and cloud systems can help companies understand where assets are, what condition they are in, and what may happen next.

The conversation also covers the role of human judgment. Supply chains require many decisions, often with consequences that spread across the system. That makes guardrails and human involvement especially important.

Key Takeaways

• Physical AI connects AI systems with data from the real world.
• Better supply chain visibility starts with accurate, current physical data.
• Real time decisions matter, but decision accuracy matters more.
• Human judgment remains important when AI affects physical operations.

Highlights

01:53 What separates physical AI from traditional AI
03:18 Why real world data collection changes the problem
07:19 Supply chain visibility and practical use cases
11:57 How quickly physical AI systems can make decisions
12:48 Why guardrails matter in supply chain automation
15:28 Robotics, distributed physical AI, and connected systems

Follow The Tech Trek for more conversations about AI, engineering, product, and technical leadership.