AI Agents, Deepfakes, and the Problem of Trust

Catching phishing attacks is only half the challenge. An AI security model also has to avoid blocking legitimate emails.Ryan Luo, Founder and CTO of Aegis AI, joins The Tech Trek to explain why his team moved beyond commercial AI model APIs and began building its own models for email security.Existing models were surprisingly capable of identifying attacks, but distinguishing suspicious activity from actual threats required more precision. Ryan explains the tradeoffs between detection accuracy, cost, and control, including why the last few percentage points of model performance can be the hardest to improve.The conversation also explores how AI coding agents are changing engineering work. Ryan shares why his engineers spend time understanding real security threats, how junior engineers can develop better judgment, and why defining clear outcomes matters more as agents take on additional implementation work.Key TakeawaysCustom models can offer greater control over accuracy and compute costs.False positives can undermine an otherwise effective security product.Engineers need domain expertise to evaluate and improve AI output.Effective agentic development starts with clear goals and constraints.Episode Highlights01:52 Why Aegis AI moved beyond commercial model APIs05:06 The difficult tradeoff between precision and recall07:08 Cost and compute advantages of custom AI models14:30 Why detection engineers must understand real attacks16:27 What makes junior engineers valuable in the AI era28:21 How AI agents are changing engineering leadershipOne Line That Stuck"They can't turn off the adversaries, but they can turn off you."Follow The Tech Trek for more conversations about engineering, AI, and building technology companies.

AI is helping healthcare teams move faster. But faster models do not always mean better patient care.Hetal Patel, VP of Healthcare Economics and Data Science at Harmony Care, joins The Tech Trek to discuss where AI is delivering value in healthcare and where human expertise remains essential.Hetal shares how her team uses AI for clinical documentation, contract analysis, coding, data migration, and predictive modeling. In one example, AI helped shorten the development time for a patient readmissions model by more than 50%.But speed creates new challenges. Clinicians need time to adopt new tools, models need careful testing, and biased data can lead to biased decisions. The conversation also explores healthcare data sharing, patient privacy, and the need for responsible AI governance.Key TakeawaysAI is reducing administrative work and helping healthcare analytics teams handle more projects.Faster model development creates new expectations for clinicians adopting those models.Human expertise and bias testing remain essential when models influence patient care.Better access to clinical data could improve predictions, but raises privacy concerns.Episode Highlights01:49 How AI scribes, agents, and coding assistants are being used04:02 Measuring ROI beyond simple time savings05:52 Cutting predictive model development time by more than 50%08:03 Why faster deployment is not always the right choice10:34 How bias enters models and why testing matters16:51 The promise and privacy challenges of connected healthcare dataOne Line That Stuck"Data is dirty. Data is biased. If you build on that, you're gonna have a biased model."Follow The Tech Trek for more conversations about AI, data, and technical leadership.

Some healthcare professionals want to brainstorm about AI. Others are reluctant to use it. And many simply want technology that helps them do their jobs.Dany Jradi, Head of Product at Abby Care, joins The Tech Trek to discuss what healthcare stakeholders actually care about when AI enters their workflows. For engineers and product teams, the technology itself can be exciting. For clinicians and operators, the outcome often matters more than how the system works.Dany explains why AI adoption remains uneven, even for tools like AI scribes that have been available for years. He also discusses why reviewing administrative documents requires a different level of scrutiny than providing clinical guidance.The conversation turns to family caregiving and the potential of daily care notes. Could AI help clinical teams identify changes in a patient's condition before those changes lead to a serious event?Key TakeawaysHealthcare stakeholders often prioritize their work over how an AI system operates.AI adoption varies widely, even among technology savvy clinicians.Administrative automation and clinical guidance require different standards.Daily caregiver notes could reveal patterns that individual events miss.Episode Highlights02:02 Why healthcare professionals respond differently to AI04:20 What builders care about versus what stakeholders need08:09 AI scribes and the uneven pace of adoption11:44 Why clinical and administrative AI have different standards15:42 Finding useful signals in daily caregiver notes20:09 How AI could identify trends between major health eventsFollow The Tech Trek for more conversations about AI, product development, and building technology that solves real problems.

What happens when AI makes writing code faster, but your engineering process still operates on two week sprints?Junaid Warwani, VP of Engineering at Brellium, joins The Tech Trek to explain how he restructured engineering at his previous company after adopting AI development tools. His team moved away from traditional sprints, gave individual engineers ownership of priorities, and shifted toward weekly planning with faster feedback.But moving faster introduced new problems. Engineers had to manage more context, product managers needed different responsibilities, and the rest of the organization had to adapt.Junaid shares what worked, where the model fell short, and what he would change as he brings those lessons into his new role.Key TakeawaysWhy traditional sprint planning started losing its value as AI accelerated development.How one engineer can own delivery while product managers focus on what comes next.Why feedback loops can become a bigger bottleneck than writing code.How hiring priorities are shifting toward communication, customer understanding, and critical thinking.Episode Highlights02:48 Why Junaid returned from parental leave ready to rethink engineering.06:22 Rethinking the relationship between product managers and engineers.10:45 Replacing traditional sprints with individual ownership and weekly priorities.14:20 Why faster feedback matters more than faster code.19:23 What Junaid now looks for when hiring engineers.27:31 Why engineering leaders need to become better at change management.A Moment Worth Pulling Out"What we're doing does not make sense anymore."Junaid describes the realization that led him to rethink his team's development process.If you're reconsidering how your engineering team works with AI, this conversation offers practical lessons from a leader who has already experimented with a different model.Follow The Tech Trek for more conversations about technology, engineering, AI, and leadership.

What actually makes a company AI native? Is it using AI across every department, building more agents, or changing how the organization works?Eswar Vandenapu, Cofounder and CTO of NovaWorks AI, shares what his team is learning as it moves from basic AI adoption toward a shared context model that shapes how decisions are made.The discussion explores why giving everyone access to AI tools is only the beginning. Eswar explains how NovaWorks is using shared context to improve collaboration, why an agent that understands its environment can be more useful than dozens of specialized agents, and how the boundaries between product and engineering are changing.We also discuss the skills engineering teams need as AI handles more coding work, including systems thinking, domain expertise, and the ability to ask better questions.What You'll LearnWhy AI adoption alone does not make a company AI native.How shared context changes collaboration and decision making.Why agent quality matters more than the number of agents.How AI is changing engineering roles, hiring, and technical skills.Episode Highlights00:40 How NovaWorks is rethinking HR and workforce management03:37 The journey from AI awareness to shared organizational context07:54 Why building more agents is not necessarily better11:07 How AI is changing the relationship between product and engineering14:41 Why systems thinking and domain knowledge matter more in engineering18:57 Learning to question AI rather than simply give it instructionsOne Line That Stuck"Quantity of the agents doesn't matter."Follow The Tech Trek for more conversations about engineering, AI, product development, and building technology companies.

A few seconds of audio can now be enough to clone a voice. But the bigger challenge is knowing who, or what, is on the other end of an interaction. That question matters when someone calls a bank, joins a job interview, or sends an AI agent to act on their behalf.Vijay Balasubramaniyan, CEO and Co Founder of Pindrop, joins The Tech Trek to discuss how identity and security are changing as deepfakes and AI agents become easier to create. He shares why Pindrop's apparent overnight success took 15 years, how fraudsters are automating account attacks, and why companies need to think beyond a single identity check.The conversation also gets into hiring. Vijay describes fake applicants, different people appearing across interview rounds, and the case for verifying identity throughout the hiring process and after onboarding. His larger point is that stronger, ongoing identity signals could also make security less intrusive for legitimate users.Key takeaways- Deepfake detection needs to keep pace with tools that can clone voices from seconds of audio.- Companies must distinguish legitimate AI agents from agents attempting fraud or unauthorized access.- Recruiting teams are increasingly part of an organization's security defenses.- Continuous identity can protect interviews, accounts, and employee access without adding constant friction.Episode highlights- 01:38 Why Pindrop's success took years to build- 03:44 Voice cloning goes from 20 hours to three seconds- 06:44 How fraudsters use AI agents to make calls at scale- 11:51 Who takes responsibility when an AI agent makes a bad decision?- 16:28 Fake candidates and the risks hiding in hiring funnels- 21:34 How continuous identity could make security feel invisibleA moment worth pulling out: “They are now part of security. They're the front end.” Vijay on recruiting teams.Follow The Tech Trek for more conversations about the decisions behind modern technology companies.

The goal is not to do less. It is to give agents the flexibility to handle more of the work customers actually need.Henry Hopcroft, Chief Product Officer at Gradial, joins The Tech Trek to discuss how building AI products is changing traditional approaches to product development, engineering, and customer delivery.Henry explains why his team has moved away from building individual agent features toward a smaller set of core capabilities. He also shares how Gradial tests agents against difficult enterprise environments, uses forward deployed engineers to learn directly from customers, and approaches product planning when AI capabilities are improving so quickly.What You'll LearnWhy adding more features can limit an AI agent's flexibility.How real customer failures shape Gradial's agent evaluations.Why forward deployed engineers play a central role in product development.How coding agents are bringing product and engineering roles closer together.Episode Highlights03:59 Why Gradial started deleting agent features.07:31 Why traditional software testing falls short for enterprise AI agents.08:35 Building a nightmare marketing stack to test agents.12:03 How forward deployed engineers improve the core product.15:56 Why great engineers increasingly need product judgment.21:55 Why long term roadmaps are becoming harder for AI companies.A Moment Worth Pulling Out"Features can be something of an anti-pattern, I think, for agentic software."Follow The Tech Trek for more conversations about AI, engineering, product strategy, and building technology companies.

Jayeeta Putatunda, AI Engineering Lead at Turing, joins The Tech Trek to explore what happens after AI generates the first draft. She explains how excessive code can create maintenance challenges, why growing pull request volumes put pressure on code reviews, and how teams can build AI review agents around their own business rules.The conversation also explores how AI is changing product development. Teams can test ideas, create prototypes, and gather customer feedback faster. But taking those ideas into production still requires testing, security, and engineering discipline.Jayeeta also discusses LLM observability, the growing importance of domain expertise, and why sharing lessons across engineering teams matters as AI tools evolve.Key TakeawaysMore AI generated code can create additional review and maintenance work.Code review agents need business context, not just generic rules.Faster prototypes still require careful testing before production.Domain expertise and human judgment remain important in AI systems.Episode Highlights00:34 Why AI generates more code and documentation than teams may need.03:46 How growing PR volumes create new review challenges.06:44 Measuring productivity beyond the number of features delivered.12:10 Managing stakeholder expectations when prototypes become easier to build.14:34 Why LLM observability matters in regulated industries.18:44 The role of domain experts in building more specialized AI models.One Line That Stuck"Who's reviewing those 10 PRs?"Follow The Tech Trek for more conversations about engineering, AI, data, and technical leadership.

AI teams can build solutions faster than many organizations can make decisions. That gap is changing how engineers work with the business and what companies need from technical leaders.Ash Pembroke, Managing Director of AI Strategy at Accordion, joins The Tech Trek to discuss the growing role of forward deployed engineers. She explains why technical depth alone is not enough and how successful teams combine engineering, product thinking, and business knowledge.The conversation covers leading lean engineering teams, finding the right mix of skills, working alongside business stakeholders, and navigating the tension between rapid AI development and slower organizational adoption.Key TakeawaysStrong forward deployed engineers combine technical expertise with curiosity and business understanding.Lean teams need complementary skills rather than one engineer who can do everything.AI agents can extend existing SaaS systems, but replacing those systems requires careful decisions.Faster development does not automatically mean faster adoption or better business decisions.Episode Highlights01:30 What makes a strong forward deployed engineer?03:18 Leading small engineering teams with overlapping responsibilities08:25 How AI agents are changing the role of traditional SaaS16:50 Moving from software engineering or product into forward deployed roles25:13 Why business decisions can become the bottleneck to AI adoption33:48 Giving business teams the ability to automate their own workflows safelyA Moment Worth Pulling Out"Nobody knows the work better than the person doing the work."Ash makes the case for giving employees the tools, access, and guardrails to improve their own workflows.Follow The Tech Trek for more conversations on AI, engineering, and technical leadership.

AI can generate answers in seconds. But what happens when the information needed to make a decision isn't available online?Sagar Kadakia, founder and CEO of Qualitate, believes reasoning is becoming abundant while real world evidence remains scarce. His company uses AI moderators to interview experts, capture their knowledge, and make that information accessible to businesses.In this episode of The Tech Trek, Sagar explains how AI can change the economics of primary research. We discuss why human knowledge still matters, how AI interviews work, and what happens when companies can gather expert insight faster and at lower cost.The conversation also explores AI's impact on jobs, why Sagar is hiring more engineers, and how early AI adoption could create lasting advantages for businesses.Key TakeawaysAI reasoning is becoming widely available, but firsthand human knowledge remains difficult to access.AI moderators can reduce the cost and complexity of collecting expert insights.AI may affect coordination roles differently from engineering, sales, and physical services.Companies adopting AI early could build advantages that compound over time.Episode Highlights00:55 | How Qualitate captures human knowledge using AI04:20 | Why reasoning is abundant but real world evidence is scarce08:20 | How AI could reshape research and consulting services12:00 | Inside an AI moderated expert interview23:00 | Which jobs could change as AI adoption increases28:00 | Why AI adoption could separate competing companiesOne Line That Stuck"Reasoning is no longer scarce."Follow The Tech Trek for more conversations about AI, engineering, technology leadership, and company building.