Tech Transformed

Artificial intelligence has spent the last several years getting smarter on screens, recommending what to watch, drafting what to write, answering what we ask. This type of progress is now spilling out into the physical world, where machines are starting to move, sense, and adapt on their own. In this episode of Tech Transformed, host John Santaferraro, Founder and Analyst at Ferraro Consulting, talks with Prith Banerjee, Senior Vice President of Innovation at Synopsys, about what happens when AI has to obey the laws of physics instead of just the patterns in a dataset.
Banerjee brings a rare vantage point to the conversation. Before Synopsys, he led HP Labs worldwide, served as group CTO at ABB and Schneider Electric, and ran engineering simulation software company ANSYS as CTO until its acquisition by Synopsys roughly a year ago - a deal that now anchors much of what he describes as Synopsys's "silicon to systems" strategy.
The Road to Physical AI
Banerjee traces AI's progress through distinct phases he's tracked across his career. It started with analytics, which was the correlation engines behind a Netflix recommendation, or the placement and routing improvements Synopsys has long applied inside its own chip design tools. Generative AI came next, giving machines the ability to produce original language, images, and video from a prompt rather than just surface existing content. Agentic AI followed close behind, handing off entire tasks, drafting a slide deck, prepping a sales call to systems that act more like assistants than tools.
Physical AI is the phase Banerjee sees unfolding now, and it's a different kind of leap. Instead of learning from words or pixels, these systems learn from real physical measurements: pressure, temperature, stress, and strain. Training that kind of intelligence takes synthetic data generated across structural, fluid, and electromagnetic physics precisely the simulation capability ANSYS brought into Synopsys.
Teaching Robots to Learn Like Humans
The shift shows up clearly in how robots are built today. A decade ago, getting a robotic arm to pick up a bottle without crushing it meant writing enormous programs, sometimes 100,000 lines of code specifying exactly how each motor should move. Physical AI throws that playbook out. Banerjee compares it to teaching a child to ride a bike: nobody narrates which pedal to push. The child watches, tries, falls, and adjusts.
Robots now learn the same way, refining their behaviour through reward and penalty as they attempt a task thousands of times. Autonomous vehicles follow the identical pattern at far greater scale, learning from millions of hours of driving footage until they recognise, for instance, that a pedestrian stepping into the street means stop. Synopsys works with autonomous vehicle and robotics companies to generate the synthetic training data that makes this kind of learning possible without requiring endless real-world testing.
Engineering the Intelligent Systems of Tomorrow
That intelligence has to run on something, and the conversation turns to what it takes to build the silicon underneath it. Chips that once held a few hundred thousand transistors now carry tens or hundreds of billions, some approaching trillions, stacked using advanced 3D and chiplet techniques. Designing them means balancing power, performance, and thermal limits simultaneously rather than simply over-engineering for safety margin, which Banerjee calls co-design.
Synopsys is tackling that complexity with what it calls agent engineers. AI systems introduced at its Converge conference that work alongside human chip designers on tasks like RTL design, test benches, and sign-off, effectively multiplying engineering capacity without multiplying headcount. The same pressure shows up at the edge, where trained AI models have to run inside a drone, car, or warehouse robot on a fraction of the power a data centre would use, with no room for cloud latency. Banerjee shares his perspective on why many AI projects struggle. He argues that the real challenge lies in balancing innovation with trust: robots operating alongside people raise questions of safety and collaboration, while autonomous systems connected to networks demand security, traceability, and clear explanations for the decisions they make.
His advice to engineering leaders is to treat this shift as organisation-wide rather than a single team's problem, from legal and marketing functions already using agentic tools to engineering teams rethinking how code gets written. The goal, as he puts it, isn't replacing people but making them capable of far more than they could manage alone. If you would find out more about this, visit Synopsys or follow Prith Banerjee on LinkedIn.
Takeaways
Evolution of AI from analytics to physical AI.
Role of synthetic data in training physical AI.
How robots learn through physical interactions.
Complexity and innovation in chip design.
Agentic AI and its applications in engineering.
Challenges of edge AI in autonomous systems.
Security, governance, and safety in physical AI.

Chapters
00:00 Introduction to AI's Evolution and Physical AI
01:15 Prith Banerjee's Career Journey and Role at Synopsys
04:19 The Phases of AI: Analytics, Generative, and Agentic
07:55 What is Physical AI and How It Learns from the Physical World
09:23 Robotics and Autonomous Learning Through Physical AI
19:41 Impact of AI on Chip Design and Complex Systems
23:29 Design Challenges of Complex, AI-Driven Chips
26:57 Edge AI and Challenges in Autonomous Devices
28:09 Implications for Organisations and Future Investments

Most conversations about artificial intelligence eventually return to the same concerns: is it going to replace my job, disrupt, and create an uncertain future? But what if we are looking at AI through the wrong lens?
On the latest episode of Tech Transformed, host Trisha Pillay sits down with Kevin Surace, an author, AI expert and technology pioneer who has spent three decades working on technologies that helped lay the foundations for products such as Siri and Alexa. His perspective is very different from the usual debate. Rather than viewing AI primarily as a threat to human work, Surace sees it as a tool that can expand what people are capable of achieving.
This shift changes the conversation from what AI might replace to what it could make possible. The discussion explores how AI can help people work faster, tackle problems that were previously out of reach and spend more time on areas where human judgment, creativity and experience still matter. For Surace, the question is not whether AI will change the way we work. That change is already happening. The more important question is what we choose to do with the capabilities it puts in our hands.
Silicon Valley Author and AI Expert
Surace holds 95 patents and is often credited as the father of the AI assistant. His work goes back to General Magic, an early-1990s Apple spinout where his team built the first digital agents and a programming language called Telescript. That early assistant technology had millions of users well before anyone was talking about chatbots.
His upcoming book, The Joy Success Cycle, grew out of a question people kept asking him: Why does he seem to enjoy his work so much? The answer, he told Pillay, comes down to sequencing. Most people wait for success before they let themselves feel joy. Surace argues it works the other way. Joy has to come first, and success follows from it.
How AI Brings Joy
Surace's favourite example is presentations. He used to spend a full week getting slides right, not because the ideas took that long to develop, but because formatting, alignment, and design details ate up all his time. Now he outlines what he wants to say, hands it to an AI tool, and gets a polished deck back within the hour.
The lesson he draws from that isn't about speed for its own sake. It's about separating the parts of a job that actually bring satisfaction from the parts that never did. Moving a font two pixels to the left never gave anyone joy, he says. Delivering an idea to a room full of people can. AI, in his view, quietly removes the first category and leaves more room for the second.
AI and the Case for More Jobs
Surace pushed back hard on the idea that AI shrinks the job market. He pointed to every major shift he's lived through from the PC, email, the smartphone, and the internet. Each one triggered the same fear, and each one ended up creating more roles than it eliminated.
He sees the same pattern playing out with software development right now. Coders using AI tools are shipping features and fixing bugs far faster than before, and instead of trimming teams, many companies are hiring more developers to keep pace with the new demand their own speed has created. The job itself has changed. Writing code line by line matters less than guiding the output and checking it. But the need for people hasn't gone away.
Rethinking Workflows With AI
One point Surace kept returning to was how companies get AI wrong when they try to automate their existing process instead of questioning why that process exists in the first place. He used car insurance claims as an example. The old approach sends an inspector to look at damage, fills out paperwork, and routes it through several people before a check gets issued. An AI-first approach skips most of that. A customer photographs the damage, uploads it, and a payment can go out the same day. This kind of redesign, he argues, is where the real gains sit. Businesses that only bolt AI onto old workflows will see modest improvements. Businesses willing to rebuild the workflow from scratch stand to cut costs dramatically and outpace competitors who don't.
Cybersecurity in the Age of AI
The conversation also turned to security, an area where Surace runs a company building biometric authentication devices. He explained that attackers rarely bother breaking into networks directly anymore, since most data sits encrypted. Instead, they use AI to generate convincing phishing emails and fake login pages designed to trick people into handing over multi-factor authentication codes. Surace believes fingerprint-based verification is where things are headed, since voices and faces can now be convincingly faked, but a fingerprint can't. He expects verified-identity badges to become common online within the next few years, giving people a way to confirm they're actually talking to the person they think they are.
What Leaders Should Do Now
When asked what businesses should do now, Surace's advice was simple: get every employee using AI tools regularly, not occasionally. He compared it to the early days of the PC, when companies had to run training sessions just to get staff comfortable with word processors and spreadsheets. Adoption took time, but the alternative was falling behind. He sees AI the same way. Leaders who wait for their teams to come around on their own risk losing ground to competitors who move faster. If you would like to learn more, visit Kevin Surace's website or follow him on Linkedln.
Takeaways
AI removes repetitive work so people can focus on higher-value tasks.
Joy drives success in the AI era.
AI will create new jobs by accelerating innovation.
An AI-first culture is key to staying competitive.
Biometric security will help counter AI-powered fraud

Chapters
00:00 Introduction to Kevin Surace and AI's Potential
01:52 Kevin's Journey and the Joy Success Cycle
04:14 AI as a Joy Maker in Daily Work
06:07 The Cycle: Joy Leads to Success
07:39 AI's Opportunity for Increased Jobs
12:15 Is AI a Technology Cycle or a Transformation?
13:50 The Ubiquity of Smartphones and AI's Role
14:39 Workforce Transformation and AI Adoption
23:06 Cybersecurity Challenges in the AI Era
26:55 Advice for CEOs and Leaders in an AI World

AI is changing the conversation around legacy modernisation, but successful transformation demands far more than powerful models and automated code generation. It requires engineering discipline, governance, and a clear-eyed view of where AI genuinely adds value and where it doesn't.
On a recent episode of Tech Transformed, host Christina Stathopoulos, founder of Dare to Data, sat down with Shodhan Sheth, Enterprise Modernisation Platform and Cloud Lead, and Alessio Ferri, Lead Software Engineer, both part of Thoughtworks' Global Legacy Modernisation Service Development Team, to unpack exactly that.
Legacy Modernisation With AI
The conversation around AI in legacy modernisation is often muddied by marketing noise. As Sheth puts it, "value and hype can coexist". Overpromising doesn't automatically mean a technology is worthless. The real question for technology leaders is whether AI meaningfully improves the cost-time-value equation for a specific problem.
This will always start with problem-solution fitness: has someone already solved a comparable challenge with AI, and does the proposed use case genuinely fit that pattern? As Sheth notes, "most things can be judged by cost, time, and value", which is a simple but effective filter for cutting through the noise.
Rethinking Legacy Modernisation
Generative AI is inherently probabilistic, while enterprise software has always relied on deterministic, predictable behaviour. Ferri unpacks this tension by separating two very different use cases: using AI to build systems, and embedding AI within operational systems.
When AI writes code, inconsistency is manageable; developers review, test and refine the output before it ships. Production systems are a different matter entirely, where unpredictable behaviour carries real operational risk. As Ferri explains, "AI in production requires different guardrails than AI for building."
The practical answer is controlled use. AI might suggest alternative products in a marketplace, for instance, while deterministic rules still guarantee that only in-stock items are ever shown. This lets AI add value within a firm, enterprise-grade constraints.
Why AI Alone Won't Modernise Legacy Systems
Technology is only part of the story. Sheth is clear that modernisation is fundamentally about change, and change is hard, especially across large enterprises with tangled, interconnected systems. Tasks that resist automation are often the hardest, like upskilling teams, explaining complex trade-offs, and winning buy-in; these cannot be solved with code alone. These human and organisational factors are routinely underestimated. Where the impact of a change is broad, he also advises either aligning teams properly across the business or breaking the change into smaller, more manageable pieces, a strategy that reduces resistance and smooths the path to adoption.
If you would like to learn more about this, visit Thoughtworks or connect with both Sheth and Ferri on LinkedIn.
Takeaways
Applying AI to modernise complex enterprise systems.
Distinguishing hype from practical AI applications.
Balancing probabilistic AI with deterministic enterprise software.
Organisational and leadership challenges in AI modernisation.
Building control, traceability, and abstractions in AI workflows.
Advice for CIOs and CTOs on AI adoption.

Chapters
00:00 Introduction to AI and Legacy Modernisation
01:18 Meet the Experts: Shodhan and Alessio
02:47 Distinguishing Hype from Value in AI
05:51 The Tension Between Probabilistic and Deterministic Systems
10:13 The Human Element in Modernisation
17:12 AI's Role in Enterprise Transformation
19:26 Distinguishing Hype from Value in AI
22:13 Probabilistic vs Deterministic Systems
27:45 People and Processes in Modernisation
29:03 Lessons in AI for Legacy Modernisation

Every customer service call now produces a measurable outcome, which is exactly why customer experience has become the place where agentic AI either proves itself or falls apart. Most enterprise AI hype centres on generation and productivity. But according to Sneha Iyer, who leads AI value delivery and analytics at Observe.ai, the real test of agentic AI is happening somewhere less flashy, which is the contact centre. In a recent episode of Tech Transformed, host Ravit Jain asked Iyer why customer experience (CX) has become the industry's default stress test and what that means for enterprises still sorting isolated chatbots from genuine AI agents. Iyer has spent nearly a decade watching CX change from manual quality assurance to a blended workforce of human and AI agents. This shift, she says, has quietly rewritten the metrics, tooling, and governance decisions every enterprise leader now has to make.
Why Customer Experience Became AI's Testing Ground
Iyer explains that CX generates more human-AI interaction data than almost any other business function. Every conversation resolves in a clear outcome, resolved or not, retained or churned, and it spans every modality a company touches, from voice to chat to email. Solve AI reliability here, and the lessons transfer everywhere else.
What's actually changed, she notes, isn't that AI got smarter; it's that AI stopped just watching and started acting. Where AI agents once flagged patterns in a transcript after the fact, they now issue refunds, verify identities, and book appointments in real time. This has split the workforce in two, namely, AI agents absorb the high-volume, low-risk interactions, while human agents get pushed toward the complex, judgment-heavy calls machines shouldn't handle alone. Old metrics like containment and deflection no longer capture that split; leaders now need blended resolution measures that account for both.
Fragmented AI Stacks
If you ask most contact centres what their AI stack looks like, you'll find what Iyer calls a "Frankenstein stack": one tool for analytics, another for telephony, another for conversational intelligence, stitched together and barely talking to each other. That setup was tolerable when each tool did one isolated job. It breaks the moment you add agentic AI, because these types of systems depend on shared context to function.
Without that context, a customer verified by an AI agent has to repeat everything to the human agent who picks up next, which turns an already frustrating support call into a worse one. Fragmented tools also make it nearly impossible to prove ROI, since every vendor tracks success differently and costs stack with each new integration.
This is the case for unified CX platforms: not as a nice-to-have, but as the only way to preserve context across a customer's full journey. Iyer points to "companion agents" as a concrete example of real-time, in-context support for human agents that shows them exactly why an AI agent escalated a call and what happened before they picked up.
Build vs. Buy
On the build-versus-buy question, Iyer highlights that building in-house looks cheap upfront and rarely stays that way. Models change every few weeks, which turns any custom build into a permanent maintenance commitment most teams underestimate, including in regulated industries like healthcare and finance, where leaders often assume building keeps data safer. In practice, established vendors have already solved the compliance groundwork, data residency, model governance, and audit trails that a custom build would take months to replicate.
Trust, she argues, isn't really about doubting AI's competence. It's the fear of the one catastrophic interaction happening at scale. That's what simulation testing and evaluation frameworks are for, stress-testing AI agents against the hardest, most ambiguous scenarios before launch, not just the easy paths, backed by governance frameworks that can roll back a deployment fast if something breaks.
Looking six to twelve months out, Iyer expects three shifts, continued consolidation around unified platforms, evaluation frameworks becoming the real competitive edge (not raw model performance), and pricing models moving from seat-based to outcome-based. Her advice to leaders is to build auditability and governance into your AI strategy now, before regulation forces the issue.
Building Customer Experience
The throughline of Iyer's conversation with Jain is that agentic AI in customer experience succeeds or fails on infrastructure and trust, not on which model you're running. Enterprises that unify their data, rethink their success metrics, and choose vendors who can prove reliability through thorough testing are the ones positioned to scale AI-driven CX without breaking it. If you would like to find out more about this visit observe.ai or connect with Sneha Iyer on LinkedIn.
Takeaways
Shift from isolated AI tools to integrated agent systems.
Importance of simulation testing and evaluation frameworks.
Benefits of unified customer experience platforms.
Build versus buy decision in enterprise AI.
Future trends in AI-powered customer experience.

Chapters
00:00 Introduction to Tech Transform and AI Evolution
02:01 Sneha Iyer's Role and Background in AI
09:35 The Shift in Customer Experience with Agentic AI
13:24 The Move Towards Unified Customer Experience Platforms
18:55 Build vs. Buy: Navigating AI Solutions in Enterprises
20:59 The Cost of AI Maintenance and Vendor Insights
22:20 Navigating Compliance in Regulated Industries
24:52 Building Trust in AI Adoption
30:00 Use Cases Driving AI Value
34:31 Future Trends in AI Customer Experience

The semiconductor industry is undergoing one of its most profound transformations in decades. Driven by the insatiable demand for compute power largely fueled by AI workloads, engineers are moving away from traditional monolithic chips and shifting toward complex multi-die designs. This shift brings a new set of challenges that conventional design and validation methods simply cannot handle.
In a recent episode of the Tech Transformed podcast, host Dana Gardner sat down with Shekhar Kapoor, Executive Director of Product Line Management at Synopsys, to explore how the growing complexity of semiconductors is changing the way engineers design and validate modern systems. From thermal management to AI-driven automation, the conversation reveals why the old way of building chips is no longer good enough and what the future looks like.
Multi-Die Design
Kapoor explains that the transition to multi-die design is no longer a matter of preference but a necessity. He attributes this shift to the relentless demand for greater compute capacity, driven largely by the rapid growth of AI.
Traditional monolithic chips are hitting hard limits. Reticle sizes are maxing out, and rising yield and cost challenges make it increasingly impractical to pack more functionality onto a single die. Multi-die designs solve this by disaggregating functionality across smaller dies, each targeting the most appropriate process technology, then integrating them into a unified, optimised package.
Leading AI systems already integrate multiple compute and I/O dies alongside large high-bandwidth memory (HBM) stacks, scaling to 3x–5x reticle-class designs and beyond. The design challenge is very different. As Kapoor puts it: "You're no longer optimising a single chip, you're optimizing a system of chips."
This requires system-level co-design from day one, spanning architecture, silicon, packaging, power delivery, and interconnect strategy simultaneously. Engineers must think in terms of System Technology Co-Optimisation (STCO), not just chip-level optimization. The design tools, methodologies, and team workflows all need to change. For engineers and technology leaders looking to explore these trade-offs, Synopsys has published a comprehensive eBook on accelerating multi-die design and innovation.
Thermal Analysis and Multi-Physics Validation
Historically, thermal, power, and electromagnetic analyses were performed as downstream validation steps once the core design was complete. In a multi-die world, that approach is no longer viable.
"Thermal management is becoming the number one issue when designing these multi-die designs. It has to be managed across a range of scales, from transistor activity to package and board level," Kapoor says.
The problem with late-stage validation is timing. By the time thermal or power integrity issues surface, the most critical decisions are already locked in floorplans, interconnect topologies established, and packaging assumptions embedded.. At that point, the only options are costly ECOs, excessive margining, or a full redesign. Industry estimates suggest over-design can lead to up to 30-35 per cent wasted silicon and hundreds of millions of dollars in optimisation loss.
The solution is a shift-left approach that embeds multiphysics analysis from the earliest stages of design. When thermal hotspots, voltage drop issues, and electromagnetic interactions are identified early, engineers can adjust partitioning and placement strategies before they become expensive problems.
This is the methodology detailed in the Synopsys ebook on Multiphysics Fusion for multi-die design, which covers how teams can build continuous multiphysics validation into their flows to avoid late-stage surprises and protect both performance and reliability.
Multiphysics Fusion and AI-Driven Chip Design
To operationalise the shift-left methodology at scale, Synopsys has introduced the concept of Multiphysics Fusion. This is the native integration of AI-powered EDA technologies with ANSYS's gold-standard multiphysics sign-off analysis capabilities.
Within the 3DIC Compiler platform, this means unifying the implementation environment with RedHawk-SC, RedHawk-SC Electrothermal, and HFSS-IC technologies. This brings IR drop, thermal, signal, and power integrity analysis directly into the design loop. The result is greater predictability, tighter correlation between in-design analysis and sign-off, and significantly fewer design iterations.
The impact on design closure times has been substantial. According to Kapoor, teams using the Multiphysics Fusion solution have seen turnaround times shrink "from weeks to days, and in some cases even hours" even for large, high-performance multi-die designs.
AI amplifies these gains further. Synopsys employs AI in two primary ways: assistive automation through its 3DSO.ai technology, which integrates multiphysics feedback into the optimization loop in real time, and agentic workflow orchestration, which becomes increasingly critical as system complexity scales toward designs incorporating hundreds or even thousands of GPUs. As Kapoor notes, at that scale, "agentic workflows could help engineers converge faster" and manage trade-offs that would otherwise be intractable. If you would like to find out more about this, download the full eBook: Multiphysics Fusion Technology for Multi-Die Designs Explained from Synopsys, which expands on each of these themes with real-world examples, design methodologies, and guidance for implementation teams. You can also connect with Shekhar Kapoor on LinkedIn.
Takeaways
Multi-die architectures and their drivers.
Challenges of traditional monolithic chips.
Importance of early multi-physics analysis.
Multiphysics fusion and its benefits.
AI's role in design automation.
Reducing time-to-market through integrated platforms.
System-level co-design.
Thermal management in 3D IC stacking.
Shift left approach in multi-physics validation.
Future trends in semiconductor design.

Chapters
00:00 Introduction to Semiconductor Complexity
02:00 The Shift to Multi-Die Designs
04:30 Challenges in Multi-Die Design
08:11 The Importance of Early Multi-Physics Analysis
10:05 Introducing Multiphysics Fusion
12:37 AI's Role in Semiconductor Design
16:37 Reducing Time to Market
19:39 Applications Beyond AI
21:12 Real-World Examples of Multi-Physics Validation
26:20 Practical Advice for Engineers