The Tech Trek

Machine learning teams are moving faster, but the hard part has not disappeared. The work is shifting from writing and debugging every line of code toward defining the right problem, setting requirements, reviewing outputs, and deciding what belongs in a durable platform.

Niels Bantilan, Chief Machine Learning Engineer at Union AI, explains how machine learning work has changed, why coding agents are accelerating prototyping, and what engineers must consider when building infrastructure that supports many teams instead of optimizing one model. He also shares how customer needs become product decisions, why machine learning roles are becoming more specialized, and why measuring AI productivity remains difficult.

Key Takeaways

• Coding agents reduce time spent on implementation, debugging, and exploration, but engineers still need judgment around architecture, quality, and business value.

• Platform teams must balance experimentation with stability by giving users freedom at the edges while protecting a reliable foundation.

• Machine learning engineering now spans a wider range of skills, from low level performance work to customer empathy, education, documentation, and developer advocacy.

• The best model for a task may depend on complexity. Smaller self hosted models can handle tightly scoped changes, while longer and more complex work may still require stronger hosted tools.

Episode Highlights

00:50 What Union AI means by an AI runtime for production

02:10 How machine learning work has changed over the past five years

10:40 The mindset shift from model building to platform engineering

15:00 Turning customer problems into reusable product capabilities

19:00 Why machine learning roles are becoming more specialized

21:50 Using coding agents through specifications, tickets, and code review

26:50 Token costs, productivity measurement, and choosing the right model

One Line That Stuck

“I’m still solving problems. It’s just the level at which I’m doing it doesn’t require me to necessarily get into the weeds of the implementation.”

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

Healthcare providers can wait 60 to 75 days to get paid, while many hospitals spend 5% to 7% of revenue on the collection process. That makes revenue cycle management more than a back office issue. It affects margins, staffing, patient experience, and access to care.

Akash Magoon, cofounder and CEO of Adonis, joins The Tech Trek to explain how agentic AI can help medical groups and hospitals automate denials, accounts receivable work, and other manual billing processes. He also shares how Adonis applies AI internally across engineering, sales, and customer success.

The conversation goes beyond automation. Akash explains why healthcare companies often win through distribution, not product quality alone, why focused solutions can create more progress than broad attempts to fix healthcare at once, and why leaders need to frame AI as a tool that helps people work at the top of their license.

Practical Takeaways

• Start with a narrow, material problem rather than trying to rebuild healthcare all at once.

• Measure AI through business outcomes, including net collection rate and cost to collect.

• Invest in marketing and distribution early, even when the product is strong.

• Build employee trust by showing how AI improves effectiveness, not only efficiency.

Approximate Highlights

00:45 How Adonis applies agentic AI to revenue cycle management

02:05 Lessons from building a second healthcare technology company

04:45 Using AI for customers and inside the company

08:55 Why healthcare progress often starts with focused swim lanes

14:25 The distribution lesson Akash carried into Adonis

20:40 How operational efficiency may improve patient access and rural healthcare

One Line That Stuck

“Healthcare ends up becoming a very humbling place to build.”

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

Most companies are not short on data. They are short on the time, cost, and coordination required to turn it into action.

Ethan Ding, co founder and CEO of TextQL, joins The Tech Trek to explain how AI agents are changing enterprise analytics. The conversation moves beyond faster dashboards into a larger shift, analysts managing fleets of agents, business teams asking far more questions, and companies finding revenue and cost opportunities that were previously too expensive to pursue.

What Technical Teams Can Take From This

• Making answers cheaper does not reduce analytics work. It increases the number of questions people ask.

• Analysts may spend less time assembling dashboards and more time managing agents, data sources, permissions, quality, and costs.

• The clearest ROI comes from decisions with direct financial outcomes, including fraud prevention, upsell opportunities, churn risk, and unused vendor spend.

• Faster analysis matters most when teams can act on valuable opportunities they previously could not afford to investigate.

• Token costs will force AI companies and buyers to reconsider where software budgets go, especially across BI tools and data platforms.

Moments Worth Hearing

00:00 Ethan explains how TextQL agents work across messy enterprise systems including Cognos, Teradata, Snowflake, Databricks, Tableau, and Power BI.

04:52 Why giving people faster answers does not create free time. It creates even more demand for analytics

07:10 How self service analytics quickly moves from asking what a number is to asking whether it matters and what to do next.

10:08 The analyst role shifts toward managing fleets of agents and tuning an insight factory for the business.

14:38 Why faster access to data can reveal valuable opportunities that were previously too expensive to investigate.

19:55 A practical way to measure analytics ROI through fraud prevention, upsell opportunities, and other direct financial outcomes.

24:18 How token costs, AI margins, and easier migrations could reshape spending on traditional BI tools.

One Line That Stuck

“It becomes much more of an operations manager job. It is a factory. It takes in tokens and churns out dashboards, reports, and recommendations.”

Follow The Tech Trek on your podcast platform, subscribe for future episodes, and share this conversation with someone rethinking how their team works with data.

Enterprise AI is easy to demonstrate. The real test begins when a promising POC meets production costs, security requirements, data movement, latency, and internal adoption.

Shimon Ben-David, CTO at WEKA, joins Amir to discuss the gap between experimenting with generative AI and operating it at scale. They explore how classical AI differs from generative AI, why production exposes problems that demos hide, and how companies with limited AI maturity can start building useful internal capability.

Practical Takeaways

• A successful POC proves that an outcome is possible. It does not prove that the system will be affordable, secure, reliable, or fast at scale.

• Enterprise AI adoption reaches across infrastructure, engineering, data, security, and business teams. It cannot be owned by one group in isolation.

• Adding more GPUs will not fix slow data access, poor utilization, weak pipelines, or an experience users do not want to use.

• External support can help, but the person or firm involved needs to stay through implementation and production, not stop at recommendations.

• Companies that are behind should begin with proven use cases, build internal experience, and quickly stop experiments that fail to show value.

Key Moments

00:00 Why moving enterprise AI into production remains difficult

01:55 The difference between classical AI and generative AI adoption

07:05 How companies can use AI without having a formal AI strategy

11:35 Why successful POCs often struggle when they reach production

17:35 Competitive pressure, AI FOMO, and the need to calculate real ROI

22:00 Why AI adoption requires cross organizational change

33:10 Where a company with limited AI maturity should begin

One Line That Stuck

“The promise is there. It is possible. You just need to do it properly.”

Subscribe to The Tech Trek for more conversations about how technical teams are building, operating, and adapting around AI, data, product, platform, and engineering execution.

AI can generate code faster, but that does not make software delivery simple. It shifts the pressure to requirements, architecture, review, and technical judgment.

Goncalo Silva, CTO at Doist, explains how AI is changing the way teams behind Todoist and Twist build software. He shares why greater individual autonomy has led to more collaboration, why deep expertise still matters, and how faster execution is reshaping product delivery, project planning, and engineering hiring.

What Leaders Can Take From This

• Faster code generation makes strong planning and clear requirements more important, not less important.

• Designers, product leaders, and engineers can work from richer prototypes, but production systems still need experienced technical judgment.

• Engineering capacity does not have to move into other functions. Teams can use it to improve reliability, performance, quality, and the amount of valuable work they ship.

• Token counts are a weak measure of progress. Doist looks at team feedback and whether projects are staying on track.

• Engineering interviews need to test architecture, decision making, curiosity, and depth, not simply whether a candidate can produce working code.

Approximate Highlights

00:00 Meet GonCalo Silva and the products behind Doist

02:00 How broadly AI is being used across Doist

04:15 Why greater autonomy has brought teams closer together

09:45 Where nontechnical coding works, and where it creates risk

17:50 How AI compressed a major refactoring effort by 20 to 30 times

25:05 Measuring AI value without counting tokens

30:20 Why faster execution requires more up front planning

34:50 How Doist changed its engineering interview process

One Line That Stuck

“We are the bottleneck. Our attention span, our ability to memorize, our ability to understand, and deep expertise.”

Follow The Tech Trek for more conversations on how technical teams are changing the way they build, hire, and operate.

AI is not just changing how engineers write code. It is changing who gets close enough to shape the work.

In this episode of The Tech Trek, Robert Stewart, CTO at Arbital Health, joins Amir to talk about how AI is bringing actuarial subject matter experts closer to product and engineering teams, especially in healthcare and risk based contracts. Robert shares how his team is pairing technically minded SMEs with software engineers, using AI tools in development, and rethinking technical hiring now that AI assisted coding is part of the job.

Practical Takeaways

• AI can reduce the distance between domain experts and engineering when the SMEs can clearly describe requirements, acceptance criteria, and edge cases.

• Pairing a subject matter expert with an experienced engineer can be more powerful than traditional pair programming because each person brings a different kind of judgment.

• Better written requirements matter more in an AI assisted workflow because tools can work directly from detailed tickets and context.

• Technical interviews may need to test how candidates use AI, not whether they can avoid it.

• Hiring teams need stronger signals around identity, environment fit, prompting skill, and how candidates respond to AI output.

Timestamped Highlights

00:00 Robert Stewart on Arbital Health, value based care, and the role of actuarial expertise in healthcare infrastructure.

03:06 Why actuarial knowledge is hard to transfer into engineering teams through normal handoffs.

04:40 How AI helps subject matter experts move closer to product and engineering work.

06:08 Why engineering fundamentals still matter, even when AI makes code easier to create.

09:55 How Arbital Health is using Cursor, Claude Code, and human review in a regulated environment.

14:52 Why more detailed Jira tickets are becoming more valuable in AI assisted development.

17:10 How AI is changing technical interviews from “you may use AI” to “you must use AI.”

22:16 What suspicious candidates, remote interviews, and fake profiles are forcing hiring teams to rethink.

One Line That Stuck

“You can judge an expert by the type of questions they ask.”

Pro Tips

• Ask candidates to share their screen during AI assisted technical interviews.

• Watch how they prompt, not just what they produce.

• Look for whether they catch strange or weak AI output.

• Use a rubric, but also evaluate whether the candidate fits the way your team actually works.

• For AI generated code, add stronger human review, especially in regulated environments.

Subscribe to The Tech Trek for more conversations on how technical teams are adapting around AI, data, product, platform, hiring, and engineering execution.

Voice AI is moving from simple call routing into work that used to require trained human agents. The harder question is what happens when those conversations involve lending, collections, servicing, compliance, and real customer risk.

In this episode of The Tech Trek, Amir Bormand speaks with Joshua March, founder and CEO of Veritus, about building AI voice agents for regulated financial services. Joshua shares why consumer lending is a demanding test case for voice AI, what makes regulated conversations different, and why the next version of the contact center may be built with much smaller teams overseeing AI systems.

Practical takeaways

AI voice agents only matter if they can actually resolve the issue. Joshua argues that users have been trained to distrust automated phone systems because most IVRs block progress instead of helping.

Regulated communication is not just about what the agent says. It also includes who can be contacted, when they can be contacted, call frequency, TCPA rules, QA, and post call compliance.

Complex voice agents require more than a prompt. Joshua talks about context engineering, state management, specialized background agents, compliance monitoring, KYC workflows, latency, and turn detection.

AI changes startup execution. Small teams with experienced people can build and ship much more than before, but that also raises the pressure to move faster.

Venture backed AI companies face a bigger bar. Joshua makes the case that higher seed valuations and larger funds increase the need for very large outcomes.

Timestamped highlights

00:00, Why Veritus is focused on AI communications for financial services and voice agents in consumer lending

02:10, Joshua’s path from Facebook apps to social customer service, messaging, bots, and now voice AI

05:05, Why AI voice agents may replace a large share of traditional call center work

07:00, Why customers have learned to fight IVRs and what changes when AI can actually solve the problem

10:00, The compliance layers around regulated voice conversations in lending, servicing, origination, and collections

14:00, Why production voice agents need context engineering, state machines, background agents, observability, and monitoring

20:30, How AI has changed startup hiring, management, productivity, and the role of experienced individual contributors

One Line That Stuck

“This isn’t a crappy IVR that’s just trying to get in my way, this is an intelligent system that can actually take actions and actually resolve my issue.”

Practical lens for technical teams

If you are building AI into customer operations, the hard part is not only getting the model to speak well. The harder work is making sure it knows what it can do, when it can act, what rules apply, how it is monitored, and when humans need to step in.

That matters even more in regulated industries, where the conversation itself is only one part of the system.

Follow The Tech Trek for more conversations on how technical teams are building, operating, and adapting around AI, data, product, platform, and engineering execution.

Joanne Chen, VP of Data and AI at SimplePractice, joins The Tech Trek to talk about what it takes to build AI data products in a regulated, sensitive domain where privacy, consistency, monitoring, and customer trust have to be designed from the start.

This conversation gets into why AI product development feels different from traditional software, how teams should think about quality control, and why not every valuable AI solution needs to be GenAI.

Practical Takeaways

• AI products need defense in depth, especially in healthcare, where privacy, confidentiality, and security cannot depend on one layer of protection.

• The core product questions still matter. What customer pain does this solve, who benefits, and what does it take to ship responsibly?

• AI changes the development life cycle because outputs are not always deterministic and quality can degrade after launch.

• Teams need monitoring, validation, and a plan for edge cases before putting AI features in front of customers.

• AI literacy is becoming part of every role involved in building, marketing, supporting, and operating software products.

Timestamped Highlights

00:00 Joanne Chen on AI data products, deterministic outputs, and safely shipping AI features

01:20 What SimplePractice does for mental health practitioners and group practices

02:30 Why healthcare AI needs multiple layers of risk protection

05:00 What makes an AI data product different from a traditional data product

08:15 Why stakeholder expectations around AI have widened so much

10:40 How AI changes the work across engineering, CS, marketing, and support

13:10 Where AI can help reduce tedious administrative work in healthcare

16:45 Why leaders need to keep their hands dirty with new AI tools

One Line That Stuck

“Keeping hands dirty is important.”

Subscribe to The Tech Trek for more conversations on how technical teams are building, operating, and adapting around AI, data, product, and engineering execution.

Kevin Haggard, Vice President of Engineering at Barracuda, joins The Tech Trek to talk about how AI is showing up across cybersecurity products, engineering workflows, team adoption, and software delivery culture. He shares how Barracuda is approaching AI with guardrails, why adoption varies across teams, and what happened when the company ran protected AI dev days for the engineering organization.

What to take from this episode

* AI adoption inside engineering teams will not be even. Some teams are already orchestrating agents from requirements to deployment, while others are still figuring out where AI fits into their day to day work.

* Guardrails matter more in security sensitive environments. Barracuda uses an AI gateway and control plane so teams can experiment without leaking data or letting agents take uncontrolled actions.

* Protected time changes behavior. Barracuda’s AI dev days gave teams three days with no meetings so they could work with the tools inside real projects instead of treating AI as a side experiment.

* The coding bottleneck may move. AI can create more code faster, but QA, testing, release safety, problem definition, and rollback mechanisms become even more visible.

* The engineer’s role is shifting from operator to orchestrator. Kevin argues that system design, review, context, and crisp instruction will become more valuable as agents take on more execution.

Key moments

00:28, What Barracuda does and how AI pairs with people in cybersecurity

03:35, Why AI adoption varies across engineering teams inside a larger organization

05:06, The need for AI guardrails, gateways, and control planes

06:20, How Barracuda ran AI dev days across the organization

08:42, The light bulb moments from product, design, and engineering teams

13:07, Why AI velocity makes existing delivery bottlenecks harder to ignore

17:00, How engineers may move from writing code to orchestrating agents and reviewing systems

One line that stuck:

“Their role is going to elevate more.”

Practical moves from Kevin’s experience

* Bring trusted partners in for training, but follow that with hands on sessions.

* Give teams protected time to use AI inside actual work, not just demos.

* Share what the advanced teams are learning so adoption does not stay isolated.

* Keep people in the loop, especially when agents are generating code, tests, or workflow changes.

* Work backward from release bottlenecks, not just coding speed.

Subscribe to The Tech Trek for more conversations on how technical teams are adapting around AI, data, platform, product, and engineering execution.

#agenticai #ai #techleadership #engineeringleadership #engineering 

AI adoption is no longer just a policy conversation. For many organizations, the bigger question is how to move faster without creating avoidable risk.

In this episode of The Tech Trek, Amir Bormand sits down with Aimee Cardwell, CIO and CISO in residence at Transcend, to talk about responsible AI deployment, the tension between speed and control, and how leaders should think about security, compliance, productivity, and customer experience as AI moves through the enterprise.

Aimee brings a rare view across the CIO, CISO, and board lens. The conversation gets into why blocking AI often backfires, how prompt redaction can help teams move faster safely, where companies should draw the line on risk, and why some teams may need to rethink old assumptions about tech debt, code ownership, and modernization.

Practical Takeaways

• Responsible AI depends on the lens. Security, compliance, business, board, and technology teams may all define it differently.

• Blocking employee AI usage can create worse outcomes. People may use shadow tools anyway, or teams may fall behind in productivity.

• Prompt redaction and enterprise agreements can give teams room to experiment while reducing exposure of sensitive data.

• Moving fast is not the same as releasing half finished customer experiences. Bad AI tools can train customers to distrust the entire interaction.

• AI may change how teams think about tech debt, refactoring, and whether some legacy systems should be rebuilt instead of patched forever.

Timestamped Highlights

00:00 Responsible AI deployment and why the definition changes by role

02:35 Aimee explains the CIO, CISO, and board perspectives on AI adoption

05:14 Why companies that block AI may create shadow usage and slower teams

06:52 Prompt redaction as a practical way to let employees experiment safely

10:40 How AI risk changes when the data exposure model is different from traditional insider theft

15:10 Why releasing poor AI customer experiences can damage trust

21:50 Using shared enterprise prompts to raise the quality of AI output across engineering teams

26:20 How AI could change the way teams approach security debt and code modernization

One Line That Stuck

“The conversation has flipped, and it is really how can I get the company to go faster.”

Pro Tips

• Start by identifying what truly makes your business defensible. Not every asset carries the same risk.

• Give employees safe paths to use AI instead of pretending they will not use it.

• Build shared prompts with engineering standards, approved tools, and company context so teams do not start from scratch every time.

• Ask whether old assumptions still hold. Some decisions made sense when changes were expensive, slow, or risky. AI may change that equation.

Subscribe to The Tech Trek for more conversations on how modern technical teams are building, hiring, operating, and adapting around AI, data, platform, product, and engineering execution.

#ai #agentic #techleadership #engineeringleadership