Click-Ins: Turning Vehicle Damage into Data with Dmitry Geyzersky

Summary

Dima Geyzersky, CTO and founder of Click–Ins, shares insights on how AI and computer vision are transforming automotive damage assessment, insurance claims, and vehicle condition evaluation. Discover the challenges, innovations, and future trends in AI-driven automotive insurance technology.

Takeaways

AI-driven damage assessment
Synthetic data for training models
Overcoming AI hallucinations with ontology and engineering
Expanding AI solutions in US and global markets
Balancing AI automation with human expertise

Soundbites

"We bring inspectors to Superman with our AI."
"Humans must be in the loop to build trust."
"Adversarial attacks are a major threat to AI."

Chapters

01:14  The Origin of Clickins: Fighting Insurance Fraud
02:19  How AI Enhances Damage Detection in Vehicles
03:00  Core Problems Addressed by Clickins
03:47  Use Cases in Insurance and Automotive Sectors
05:47  Dima’s Focus as CTO: Synthetic Data and Deep Tech
07:19  Overcoming AI Hallucinations with Ontology and Engineering
09:32  Combining Knowledge and Prediction in AI Models
09:54  Ideal Partners for Clickins’ Solutions
12:15  Main Challenges in Growing Clickins
14:01  Avoiding AI Hype and Building Trust
17:58  Growth Targets and Market Expansion
20:02  Training on Synthetic Data for Privacy and Trust
21:25  Emerging Trends in AI and Computer Vision
27:01  Future Outlook and Industry Impact

Video

https://youtu.be/E2BZ7RvhDjc

Transcript

SPEAKER_01 0:20

Welcome to AI Speed, the show where AI-powered companies talk about what actually works in the market right now. Business doesn't move at internet speed anymore, it moves at AI speed, and the people who figure out how to turn models into money will own the next decade. I'm Evan J. Chaulfin, founder of Luxhammer and growth partner to high-performing brands. Today, I'm thrilled to be joined by Dima Gaisowski, CTO and founder of ClickIns Limited. Dimitri is at the cutting edge of applying AI, computer vision, and deep learning to revolutionize the automotive insurance claims process. His work is transforming how damage is assessed, bringing unprecedented efficiency, accuracy, and standardization to a complex industry. Dima, thanks for being here.

SPEAKER_00 1:15

Great to be here. I'm excited to chat with you.

SPEAKER_01 1:18

Absolutely. So clickins is leveraging advanced AI really to tackle automotive damage assessment. What was the pivotal insight or industry pain point that led you to identify the specific need and inspire you to found Clickins?

SPEAKER_00 1:35

So we started as an insurance anti-fraud solution. So the insurance industry was riddled with fraud. So we started combating fraud by using our technology. We are not AI-first company, so we we use AI, but we use a lot of precise engineering disciplines augmented with AI to make sure that we have a robust and reliable solution which is free of hallucinations that can be used in as an evidence in court and in and in the automotive space. So when a picture of a car is taken in the automotive space, it is considered that you see everything and you document everything. So we bring this ability to see things in images and to spot any damages and augment and make every inspector a Superman with our tool. So we don't replace humans, we allow them to make their work more efficient, find more damages, properly document, better document vehicle conditions.

SPEAKER_01 2:51

Yeah, that's fantastic. And so, what's the core problem you're solving for insurance carriers, repair shops, and ultimately vehicle owners today?

SPEAKER_00 3:01

So the insurance and the and uh the rest are completely different domains. In insurance, we solve a few problems. Uh first, starting with the underwriting. During the underwriting, most of the insurance either don't take pictures of a car or have pictures but don't analyze them automatically. So they only revert to these images when the claim arrives. So it's uh quite late. So for in the underwriting, we assess the vehicle condition and properly document it. So the insurer knows exactly what is the status of the vehicle that they are underwriting. And then the next step is claim. If claim arrives or when claim arrives, then we can first detect any damage, assess any damaged panels on a very high level, or and provide the condition of the damage and even compare it to the underwriting to make sure that only the new damage, only the incremental damage is claimed and is paid for. And uh there are other like use cases like first notification of loss, FNOL, and others that uh that our system is used across insurance. In the automotive space, the use cases start from uh from uh vehicle um vehicle remarketing or evaluating the vehicle conditions let's say in used cars, uh auctions, let's say online sale or auctions, they use our system to uh to assess images uploaded by customers, and we do this in near real time. So they have a very good uh vehicle condition report that they're used in the online. It is uh used by uh by automotive companies to avoid the arbitration or to help the in arbitration cases. In rentals, it is used to compare the drop-off and the uh and the um pickup reports. So we automatically capture vehicle condition during the pickup, and when the car is returned, we we can compare and see if there is an incremental damage that the customer must be charged for. So the there are plenty of use cases, not to mention the homeland security use case where where our system is also used to to help uh uh law enforcement and the government agencies. Yeah.

SPEAKER_01 5:41

Wow, that's fantastic. And so where are you focusing most of your energy right now as CTO and founder of ClickENts?

SPEAKER_00 5:49

So as a CTO, I am in charge of the research and technology. So as a CTO, I have to work on all the all the pipeline, all the spectrum of problems that RD issues we have. So uh our company is a deep tech, so we combine PD modeling, AI, photogrammetry, computer vision, and a lot of different disciplines. So you need to orchestrate everything to make sure that everything works together as a tandem. So today I must admit, most of my efforts go to the combination between PD and deep learning. Since we are the only company in our space that uses synthetic data for training our models, and we are absolutely ethical AI, so we don't use customers' data for training our systems. So there is a lot of work and research that goes into this direction. So making our synthetic data efficient and make it very close or even sometimes outperform real-world data. Okay, so this completely eliminates human annotation needs and a man in the middle. So all the processes absolutely annotated, automatically annotated. And another part of the research where I spend most of my work is uh combining the worlds of ontology and AI to provide visual reasoning for uh visual models, so to make sure that we can get over the AI hallucinations, which are very rampant today.

SPEAKER_01 7:23

Yeah, absolutely. Yeah, that uh how do you um how do you get over those hallucinations? What is the solution to that?

SPEAKER_00 7:31

I would say, you know, being for quite a lot of time in in technology and particularly in intelligence, I usually love, used to quote Lao Tzu, those who have knowledge don't predict, and those who predict don't have knowledge. So the idea is to move from prediction to the knowledge. And how do you move to the knowledge? So knowledge is gained by uh systematically applying uh engineering uh disciplines, not AI, the things like 3D modeling, uh reasoning through ontology, the things that you know. Ontology, this is one of my expertise. Ontology is uh the theory of existence, the things that you know that exist. Okay, so you can define entities and the context and define the world as you know. And then when you use this knowledge in the process of AI reasoning, you can validate different AI predictions and provide the reasoner that will eliminate most of the hallucinations. I would not lead you on that saying that we can remove all the hallucinations. It is it's quite a tough, okay, tough problem. But I think we are doing well, and I think we're doing a good job by reducing these hallucinations to a minimum. And we're not just saying that, okay, this is a damaged car. So if you ask any LLM to assess vehicle condition, it will tell you, okay, this fender is badly damaged, it has a dent or whatever, okay, but instead we will tell you that the fender is or the front fender is damaged, the dent size is 10 square centimeters, it is located five centimeters to the edge of the panel, it is a minor or severe damage, and we will produce uh we will project it on a 3D to make sure that it's not part of the vehicle geometry. Okay, so we try to combine a lot of disciplines to know, not to predict. So this is the one of the pillars of the visual intelligence, just not to not to look, but to see.

SPEAKER_01 9:46

Absolutely. And so look at the broader picture for a little bit. Who do you serve best? What's the ideal insurance carrier or automotive partner profile that benefits most from Clickins solutions?

SPEAKER_00 9:59

I think that most of the insurance companies uh are our potential customers, mostly those that uh have motor or sell motor insurance, because the process starts with underwriting customers, where our system today is widely used across the globe. So the users they don't have to install anything, they don't have to drive in in some facility to be inspected. They just get a link from the company and they start a very seamless and fast process of taking a few pictures, even not even a video. It's just a few shots of the camera angles of the car. So we need to see the entire car. And uh then the the detailed condition report is sent to the insurance company with the grade of the vehicle, which is uh very accurate. It was proved to be very accurate, and they know exactly how and uh what they underwrite. So uh the second step, when a claim arrives, our system is used to assess the claim. We don't settle the claim, we partner with companies that can provide cost estimation. This is not our business. So we assess vehicle condition. We are as a smart scanner running on a mobile device. And uh, with our information, any cost estimation company can write a report, estimate report, and then insurers can settle a claim very fast. And it can be used in the uh FNOL scenarios for insurers. And in the automotive space, uh car rental companies use our system on a pickup point and in a drop-off. Our system can automatically detect incremental damage caused by a customer during the rental period. Auctions can assess vehicle condition and provide a comprehensive vehicle condition to their buyers that can put better bids on the vehicles in online marketplaces can use vehicle condition to properly or adequately describe the vehicle when it is sold to avoid arbitration, and a lot of a lot of plenty of other use cases that I didn't cover yet.

SPEAKER_01 12:14

Yeah, absolutely. And so, what do you think has been your biggest challenge in growing click-ins this past year?

SPEAKER_00 12:21

And the biggest challenge, uh, we have been uh in in deep RD for uh for I would say 10 years, because the domain is very challenging. This is a very tough domain. You know, it sounds like a simple task, but it is an absolutely complicated task to uh distinguish between a scratch and, for instance, a dirt or bird droppings. How you assess vehicle condition by using only single shots. Even our eyes, a human, uses a stereo pair of two cameras in our eyes to assess uh uh objects. And a single shot, it's a 2D image, it has only one camera. So that's the main challenge to make sure that we can adequately assess the vehicle based on images, not using any hardware or video, and reduce the hallucinations of AI to the minimum to make sure that our reports are reliable and accurate, and of course, trying to spread the word of clickens and get to the customers across the globe. So we already work with uh tier one companies in the US and Europe, uh, showing very, very good results for them. And uh that's our main challenge currently is uh is expanding the clickens expansion to other markets and other geographies. Uh with a new company that we established headquarters in the US. Uh so now we are continuing our expansion in the US market. So those are, I would say, business development is currently our main challenge.

SPEAKER_01 14:08

Yeah, definitely. What would you say makes it difficult to avoid, you know, sort of the AI hype trap that a lot of people fall into and focus on the ROI when presenting the AI-driven damage assessment solution to potential clients? Because uh I know that sometimes people get into the hype of AI but don't really recognize what the solution really is.

SPEAKER_00 14:39

That's an excellent question. I'm frequently asked by investors and uh and potential investors, okay, uh, about uh this AI hype. So what happens if a new company rises or a few founders just plug some you know Chat GPT or Cloud AI into their uh into their UI, nice UI, and uh start selling uh vehicle inspection solution. So I usually say that there is no alternative to an expert system. And uh the 10 years of RD and deep research cannot be easily replaced. So usually I encourage the people to go and try that, just to see that the current state of I, or particularly LLMs or VLMs, is not there. Okay, you cannot rely on that yet. It is good, absolutely good in detecting classifying objects, it's very good in providing some general reasoning and writing text, working with language, but when it comes to the trust, we are still not there. And you probably heard about a recent case with uh uh with one of the big rental companies in the US that uh enabled a fully automated AI procedure without even the way to complain or to raise the arbitration. So so this fully automated AI process caused a lot of buzz and a lot of discussion, raised the discussion in the industry. So how reliable AI is and how to build trust with the companies. So I would say we build the trust with our customers by empowering the human inspector. We believe that the human must be in the loop, and we make a regular inspector as a professional adjuster on steroids with our tool. So uh even a junior inspector becomes a superman when they have a tool that uh works or operates as a safeguard. So we can detect, make sure that he that no damages remain undetected or under the cover. So we help them do their work much better. And all this hype around AI, yes, it's all about automation. Insurance industry, it's uh mostly undisrupted one. And I don't think that you can disrupt insurance, but you can introduce automation. So when you digitalize processes and when when you automate processes and insurance, it's uh it's worth a lot. And it may it means a lot to them and it helps them achieve their goals. So don't try, don't think that you can that you can deploy some AI, magic AI, that will uh on a sudden replace uh a very challenging and complicated insurance workflows. But you can definitely automate and help them make better decisions.

SPEAKER_01 17:42

Yeah, absolutely. I definitely think human power plus AI really is the best way to go. I've always uh felt that way. Looking at, you were talking about expanding in the US and other markets around the globe, their companies. What are your main growth targets for this year? What are you looking at in terms of growth?

SPEAKER_00 18:05

This year we are focusing mostly on insurance. On the insurance domain, we are currently negotiating with uh with biggest US carriers. We have uh successfully deployed our solution in Europe in the insurance companies in Europe. Actually, we started this a few years ago, and we are we continue our expansion on the US on the European market as well. As you know, we had to comply with uh GDPR and other data privacy regulations. So we had to invest into this space. The company was also passed the ISO certification, so to make sure that we can provide a reliable and secure solution with data privacy in mind. So today we have a very comprehensive and fully compliant solution to insurers, not only in Europe but also in the US. So I would say our current focus is mainly on insurance, but uh we also work with our tier one companies and customers in the automotive space and working on a big uh account in in the marketplaces and auctions as well. I think that some of the big news will uh will be published soon. I hope so. Okay, some of the big names and big titles. But yeah, we work in these domains and insurance, as I said, requires a lot of trust. And we do this by encouraging them to test our solution and to see. And um I think one of the key uh key selling points is uh the use of ethical AI. We never use the customers' data, we use proprietary data that we render and uh simulate internally, so-called synthetic data. So all of our models are trained mostly on synthetic data, which makes and builds trust with insurers. Because uh, you know what it means to get data from insurance company for training. It's uh it's a mission impossible because uh exactly it is subject to data privacy, it's subject to to any possible law and regulation. So, in the most cases, you will die young until you will get uh the first batch of data. And our solution was initially designed to rely on synthetic data. So that's why I think that insurance companies will embrace our technology and and they love it, because you can't pre-trained. So the POC can be as fast as five minutes running a test and seeing how it works. No humans annotating the damages behind the scene. You know, there are companies that sell semi-automated processes by having people in Bangladesh or in India that just look at the images, annotate damages. Okay? This is not scalable, and it does not build trust with insurance companies. So we provide a near real-time report, so you see that nothing can be done, nothing is manual, everything is uh fully automated, and scalable. That's very important. Absolutely.

SPEAKER_01 21:17

So, what trends are you seeing in AI, computer vision, or automotive insurance industry excite you the most, such as generative AI for claims processing, predictive analytics for fraud detection, or even the rise of autonomous vehicles?

SPEAKER_00 21:34

So everything that you mentioned will be inspired as already affected by AI, inspired by AI. And the autonomous vehicles uh is definitely built on AI advancements. So you cannot build any vision system without proper AI. But um fraud, predictive analytics, you you see now the cloud security, what it does to security companies. So you It it can detect vulnerabilities and zero-day exploits much faster than the red red team, uh red teams of uh security companies. Okay, so the you see what it can do to this uh industry. So the and in in the vision space, I think that these models made a huge progress and they need to transition from the language models to visual models because uh today it is still b it still relies on the language. So every image is interpreted into the text, text embeddings which are analyzed, but it's not a true vision system yet. Okay, so once we add physical attributes of the real world, and once the AI starts understanding the physics and optical principles, okay, behind the image, then we will see probably the next leap, the next uh step in the evolution of those AI systems. Today, it's more about prediction, it is more about predicting the scene or auto captioning the scene. Is down it is reduced to the language, to the text, to the textual presentation. And generative AI made a huge progress. Look at uh Nana Banana, what it did to the industry, and the latest version also added a special, spatial accuracy. So the the objects you can already have some kind of a control over the scene, okay, and you see less hallucinations in the in the newer models. But still, I can tell you that we are we are constantly monitoring the space. We are testing in each and every new architecture that's released to the market, but still nothing can replace the engineering process because our synthetic data is uh rendered from a 3D model using pure engineering, no AI, no uh no generative AI. We do use generative AI uh to generate some, I would say, unreal scenarios because it's very hard to find an image of uh let's say Ferrari with uh expanded uh suspension or with you know four four ten pipes or uh exhausts and in let's say uh green color. Okay, so these kind of weird scenarios they are very important. They are crucial to train AI to deal with adversarial scenes, because otherwise the model can be surprised when it sees some unreal pattern. So, in this, in this regard, using generative AI, and we do use this for very bizarre scenarios, it helps you train the model on adversarial things to make sure that when it really encounters a problem, it will not fail. So, for instance, in auto vehicles, if the model was trained on pedestrians, on uh stones, on animals, okay, on trees, just imagine you have some uh unexpected object on the road, and the model fails to detect it or to classify it. So you we train our models to be robust, to and to be stable and uh to react adequately to any adversarial attacks. By the way, today we don't hear a lot of discussions about adversarial attacks, but this is, I think, one of the biggest threats on AI. Because if the model is trained and the new generative AI is very advanced, it's just a matter of time until those models will be attacked and those they will be exploited. Because if you build processes mainly on AI without a human in the loop, and you rely on the AI, anyone will be able to, let's say, generate a vehicle image with a severe damage, send it to an insurance company, and they will assess the damage and pay the claim. So, and you can also, if you understand how these vision models work, you don't even have to manipulate, you can manipulate the image, add a certain filter to the image, and make sure that these AI models will tag this image as a severely damaged car or total lost car. This absolutely happens. But again, in in the insurance space in claims, it will lead to a bogus claim, okay? So it will be, it's only the money. But what happens if the autonomous vehicle encounters some object that is manipulated and will trick it into a sudden stop or or uh even acceleration? I don't know. It can be a disaster. So a lot of things going on, and there are a lot of threats. So you have new capabilities, new uh generation AI, but on the same side, you have a lot of threats imposed by this, and a lot of bad people there's a lot of exploitatives.

SPEAKER_01 27:20

A lot of problems to solve uh in uh in all those areas. So I'm excited to see the future of all those areas and people uh tackling each of those issues, uh, including including you uh in the industry as well. Well, that's it for today's episode of AI Speed. A huge thank you to Dima Gazersky for sharing his invaluable insights looking at how ClickIns is helping the automotive industry turn AI models into money and for navigating the cutting edge of computer vision and deep learning. If you're building or leading an AI native company or a service business that uses AI under the hood and you care about revenue adoption and market share, please make sure to subscribe to AI Speed. Learn how the best AI operators ship fast, sell smarter, and stay ahead. Thanks for listening. Until next time, keep building, keep selling, and keep moving at AI speed.