Andrei Kulik（XENOM 聯合創辦人）訪談系列

Andrei Kulik is Co-Founder and CEO of XENOM, bringing experience in on-device AI, distributed systems, and machine learning to industrial operations. He previously co-founded AIMatter, which was acquired by Google, and worked on AI systems for consumer devices at Google. His career has focused on making sophisticated AI work reliably under the practical constraints of real-world hardware and environments.

XENOM develops an industrial AI platform that combines software with smart protective equipment to give operators a real-time view of their worksites. Its system brings together positioning, workforce communications, environmental sensing, and safety data. In this interview, Kulik discusses edge AI, smart helmets, human oversight, worker privacy, and the path from industrial monitoring toward predictive decision support.

You previously co-founded AIMatter, which was acquired by Google, and spent years developing AI systems designed to run directly on consumer devices. What inspired you to co-found XENOM, and how did your experience at Google shape your vision for bringing AI into industrial environments?

After the acquisition, I spent years working on one specific problem: how do you make sophisticated AI run reliably on a device that fits in your pocket, with limited battery, unpredictable connectivity, and real-world conditions that no lab can fully simulate. We shipped that work across phones, glasses, earbuds — billions of devices. At some point I started asking myself where else this architecture could matter. Consumer devices are convenient, but industrial sites are where the cost of a wrong decision — or no decision at all — is measured in human lives and millions in lost productivity. Most industrial sites still have no real-time picture of what’s happening. That gap felt like exactly the right problem for the technology we’d spent years building.

XENOM combines AI-powered software with smart protective equipment to provide real-time visibility into industrial operations. What are the biggest blind spots in traditional industrial safety systems, and how does your platform address challenges that conventional monitoring technologies struggle to solve?

The fundamental problem is that most systems were designed to record what happened, not to prevent it happening. A worker’s location updates every few minutes. Timesheets are filled in at the end of the shift. Safety violations get logged after an incident report is filed. By the time a manager sees the data, the shift is over or the accident has already occurred. The second blind spot is fragmentation — a company might have a gas monitoring system, a separate access control system, a paper-based PPE checklist, and a manual timesheet process, none of which talk to each other. XENOM replaces all of that with a single continuous stream of data from the site, processed in real time, that gives a manager one coherent picture instead of five disconnected ones.

Much of your career has focused on optimizing machine learning and LLM inference for resource-constrained devices. How are you applying these principles to industrial environments, and what are the technical advantages of processing AI workloads at the edge rather than relying on cloud infrastructure?

Working with resource-constrained devices teaches one discipline: every millisecond, every milliwatt, and every byte has to be justified. That is exactly the mindset an industrial helmet needs, because it is worn for a full shift and has to behave as consistently at hour eleven as at hour one. We split the work deliberately: time-critical, data-heavy tasks like processing the voice stream and sensor signals or detecting falls and impacts run on the NPU in the helmet, using models compressed and tuned for that hardware, while the site-wide picture and deeper analytics are built on the platform side.

Edge processing gives us four practical advantages. The first is speed: the device reacts to an event immediately, without a round trip to a server. The second is reliability: a lean on-device model behaves predictably, while every extra link in the chain is another point of failure. The third is network load: hundreds of helmets can’t continuously stream raw audio and sensor data, so each one sends the result of the analysis instead. The fourth is battery life: transmitting is one of the most power-hungry things a device does, so processing locally and sending less extends working time.

XENOM has developed smart helmets that integrate positioning, communication, environmental sensing, and AI capabilities. What were the most difficult engineering challenges in bringing these technologies together, particularly in environments where connectivity, battery life, and reliability can be unpredictable?

The hardest problem was not any single technology — it was making all of them coexist reliably in one device in conditions that would destroy most consumer electronics. Industrial sites are dirty, physically demanding, and thermally extreme. The helmet gets dropped, exposed to rain, worn for twelve-hour shifts. Battery life becomes a real constraint when you’re running positioning, sensors, communications, and AI inference simultaneously. And then there’s the connectivity architecture: we needed seamless handoff between RTK satellite positioning outdoors and our proprietary UWB mesh network indoors, with no gap in accuracy when a worker moves between the two. Getting that transition to happen automatically, reliably, in real time, was one of the more technically demanding things we’ve built.

Your team has developed a voice AI model capable of detecting signs of alcohol impairment through speech, reporting performance that exceeds previous academic benchmarks. What technical breakthroughs made this possible, and how do you validate the system’s reliability across different languages, accents, background noise levels, and real-world working conditions?

The key insight was that intoxication leaves a detectable signature in the micro-patterns of speech — not the words, but the acoustic properties: timing, pitch variation, articulation precision. We built a model that analyzes these patterns in the background during normal voice communications, without requiring a dedicated test or any action from the worker. The 86.7% accuracy figure came from rigorous benchmark testing against published academic datasets. For real-world validation, we test across languages, accents, and background noise conditions representative of actual industrial sites — heavy machinery, wind, multiple simultaneous speakers. The model had to be robust enough to work in conditions that are nothing like a controlled study. That took significant training data diversity and iteration.

AI systems designed to identify worker impairment or fatigue introduce significant questions around false positives and accountability. How does XENOM balance early risk detection with the need for human verification, and what safeguards prevent an inaccurate AI assessment from unfairly affecting a worker?

This is something we thought about very deliberately from the beginning. The system is designed to flag a concern to a supervisor, not to take autonomous action. An AI assessment of possible impairment triggers a human check — the supervisor can observe the worker directly, ask questions, or request a physical test if warranted. The AI does not remove someone from the site. It gives a human the information they need to make a judgment call faster than they otherwise would. We also track false positive rates in deployment and continuously improve the model. The goal is to reduce the number of incidents that happen because nobody noticed — not to create a system that penalizes workers on the basis of an algorithm alone.

Before founding XENOM, you served as CTO of OneSoil, where you applied computer vision and machine learning to agricultural monitoring and productivity analysis. What lessons from deploying AI in agriculture have proven most valuable when building intelligent systems for construction sites, mines, ports, and energy facilities?

Deploying AI in the physical world is a fundamentally different challenge from deploying it in software. You’re dealing with variability, equipment that doesn’t behave the same way twice, and operators who have been doing their jobs for decades and are deeply skeptical of anything that tells them to do things differently. Industrial sites have the same dynamic. The technology has to earn trust by being reliably right more often than people expect, and it has to be simple enough that someone managing a crew of fifty can actually use it without training. The other lesson: data quality in the field is almost always worse than you planned for. You have to build systems that are robust to noise, gaps, and edge cases that never appear in a clean dataset.

Industrial AI platforms can generate detailed information about worker locations, activities, communications, and productivity. How do you approach data privacy, employee consent, and the potential for workplace surveillance while ensuring that these technologies genuinely improve safety and operational efficiency?

We think about this in terms of purpose limitation. The data XENOM collects is for safety and operational visibility — it is not a productivity surveillance tool, and it should not be used as one. We work with operators to define clear data governance policies before deployment: what is collected, how long it is retained, who has access, and under what circumstances it can be used in a personnel decision. Worker location data is aggregated and anonymized for productivity analysis — individual tracking is available to safety supervisors for emergency response, not to line managers for performance evaluation. As the EU AI Act extends to industrial AI systems, we expect regulatory frameworks to formalize a lot of this. We’d rather be ahead of that than scrambling to comply.

Moving AI from controlled laboratory environments into safety-critical industrial operations presents challenges that software-only AI companies rarely encounter. What have been the biggest surprises during XENOM’s real-world deployments, and how do you measure whether the technology is delivering meaningful improvements in accident prevention, operational efficiency, and workforce productivity?

The biggest surprise is how much value comes from simply having an accurate picture of where people are and how time is being used. We expected safety to be the primary conversation with operators. What we found is that productivity visibility — understanding where shifts are losing time, why certain crews outperform others, where the bottlenecks are — often generates faster ROI and creates internal champions for the platform much more quickly. The second surprise is how much variation there is between industries and even between sites within the same industry. Construction in Germany operates very differently from construction in the UAE. You cannot build one product and assume it fits everywhere without significant iteration.

Looking ahead, XENOM envisions an Industrial AI Copilot capable of moving beyond monitoring to predicting risks and recommending operational decisions. How do you see physical AI evolving over the next five years, and what technological breakthroughs are still needed before industrial AI systems can operate with greater autonomy while maintaining human oversight?

The current state of industrial AI is mostly visibility — knowing what is happening. The next stage is prediction: identifying patterns that precede an incident or a productivity loss before either occurs. That requires accumulating structured, labeled data at scale over time, which is exactly what XENOM is doing from day one. The stage after that is recommendation: the system not only tells you what is likely to happen, but suggests what to do about it — how to redeploy people, when to pause an operation, which crew configuration produces the best outcomes on a given task. What we’re building toward with the Industrial AI Copilot is that third stage. The technological breakthrough it requires is not primarily a modeling advance — it’s a data problem. You need enough high-quality operational data from enough diverse sites to train a model that generalizes reliably. The companies that start collecting that data now will have an advantage that is very difficult to replicate later.

Thank you for the great interview, readers who wish to explore edge AI, smart protective equipment, and real-time industrial safety monitoring should visit XENOM.