Transcript for: #538 - Why did two clinical AI founders bet on the paperwork instead? | Shaju Puthussery (CEO, Lightspun) & Deepak Ramaswamy (CTO)

This is a transcript of Slice of Healthcare #538 with Shaju Puthussery, CEO of LightSpun, and Deepak Ramaswamy, CTO of LightSpun. Please note that the transcript has been lightly edited for readability and may contain errors. Here are some useful links:

Here are the loose chapters in the conversation. Click a link to jump to that part of the transcript.

The following is a conversation with Shaju Puthussery and Deepak Ramaswamy, co-founders of LightSpun, an AI native operating system for insurance processing. Shaju spent his career inside dental insurance and was the first employee at Overjet. Deepak built quantitative systems at Interactive Brokers and GMO before co-founding Overjet. At LightSpun they made a deliberate choice: leave clinical judgment to clinicians, and rebuild the administrative machinery around it.

This is the Slice of Healthcare Podcast. And now, here are Shaju Puthussery and Deepak Ramaswamy.

Jared S. Taylor
(00:00:00) When I was doing my research, I kept coming back to how unusual your backgrounds are. Few founders have what you two have. One of you is a theoretical physicist. The other is an MIT trained computational scientist, which is wild. You both come from worlds built on finding the simple structure hiding under a messy phenomenon. So when you first really stared at dental claims adjudication, what was the underlying structure you saw that people who have been in the industry for decades had stopped being able to see?

Shaju Puthussery
(00:00:38) I will go first, Jared. It is great to be on the podcast. Before I answer, a little on my background. I have spent my career in dental insurance. I worked for a large enterprise and was fortunate to get firsthand experience of how insurance actually works. A lot of that experience looks archaic today, but at the time it was moving files by hand, manual processes, people pushing paper, working with fax machines to get information, slowly moving into the digital world.

(00:01:15) Adjudication has been manual, and it continues to be manual even in 2026. There is automation at a certain level, but a lot of pieces are cobbled together to produce that final adjudication, whether it is intake, paper, humans calling in, or someone faxing a claim. There is far more EDI adoption now, so the insurance industry has come a long way.

(00:01:50) What we have built at LightSpun is the next generation core operating system for insurance processing. We began with dental, where we have the most expertise, and we are expanding into vision and other ancillary benefits. What we have really done is replace the back end engine with an agentic AI infrastructure. More automation, less human intervention, much faster processing, cheaper for the carriers, and a lot more access for providers. That is the problem we are solving, and that is what we are excited about.

Jared S. Taylor
(00:02:26) Thank you for that opening. I want to kick it over to Deepak. Before dental, you spent years at Interactive Brokers and GMO, where a trade can clear in milliseconds and pricing is basically a solved science. Then you look at a dental claim that can take thirty days to settle. Did you see a clearing and settlement problem that finance figured out decades ago? And what does a markets background let you see in healthcare that a healthcare IT lifer cannot?

Deepak Ramaswamy
(00:03:07) That is a great question, Jared. My background at those companies reflects my training, which was on the quantitative side, building quantitative models. When I look at insurance administration, claims processing, and provider credentialing, an area you are very familiar with, I see similar problems.

(00:03:35) At an enterprise level the issues are the same. You have to transfer files. Compliance is an essential requirement. You have to be able to audit those files. So there is a lot of commonality, even if the specific compliance standards differ between insurance and finance. In both cases there is heavy adherence to process, so that you do not end up paying the wrong claim for the wrong member or patient.

(00:04:15) Where insurance could really benefit from what finance has figured out is exactly what you mentioned. Real time processing. Analytical frameworks. How do you extract the maximum efficiency out of a system. It has its own set of challenges, and not every solution transports cleanly. But to Shaju's point, a lot of what happens in insurance is deterministic and some of it requires human judgment. For the deterministic pieces, how do you make them go faster? How do you get a response back inside whatever the SLA is? A lot of it is that.

(00:05:00) The other part we are excited about is the human decision making. How do we make that quicker? Insurance comes with its own set of interesting problems that are unique to it. So, long story short, if you bring your problem solving skills you can apply them to other domains, compliance is the common theme at the enterprise level, and where insurance can learn from finance is fast transaction processing and resilient data. That is not to say people have not figured this out, but I think there are lessons to be learned there.

Jared S. Taylor
(00:05:32) This next one is for both of you, whoever wants to jump in first. At Overjet, the whole thesis was that AI should get into the clinical decision. Read the X-ray, flag the diagnosis. At LightSpun, and correct me if I am wrong, the thesis is almost the opposite. Keep AI out of the clinical judgment and let it run the machinery around it. I really like that. What made you build it that way?

Shaju Puthussery
(00:06:00) I can take the first pass and Deepak can follow. Deepak was a co-founder at Overjet and I was the first employee. We came from the same industry, and there the AI was applied to clinical diagnosis from X-rays. But what we noticed while solving that problem was a deeper, more fundamental problem on the administrative side. How do you handle a claim, intake it, analyze it, pay it? How do you run clinical review, as opposed to what happens inside the clinical review? That gave us the insight to go solve the core of the infrastructure rather than the diagnosis itself.

(00:06:50) That is what opened up LightSpun, where we use AI for administrative tasks and repeated tasks. That was in early 2022, and the models we were using then were classical models. Then LLMs came and we had to reinvent, adapt, and change. We have patented a number of the models we use for administrative AI, and we keep reinventing as the intelligence world changes very rapidly. We keep infusing that intelligence into the workflows so we can process claims faster, cheaper, and more accurately. Deepak has much more insight on the technical side.

Deepak Ramaswamy
(00:07:48) I agree with Shaju. I do not think this is an either or. You need AI in all of these places. But if you cannot get the claim to clinical review because the right documents were not submitted, or because the record the provider sent does not match, it does not do you any good to have world class AI in clinical review. You still have not solved the plumbing problem.

(00:08:20) That is where we have focused. Curating the data, making sure the data is accurate, making sure the administrative tasks are automated. The last thing I will say is that Overjet seems like a long time ago now, and the models we used then are very different from the models today. Even in our own brief existence, we started off building transformer based models and fine tuning them for our applications, that went to a hybrid with LLMs, and now it is agentic AI. It is changing rapidly.

(00:09:00) The place where we draw a line with our AI efforts is that we always have a human in the review as a first class principle in our product and our features. Whether it is reading a document from a claim and asking whether this looks okay, or going to fifteen different sources to match disparate records and asking whether that is okay, a human has to sign off on it. At the end of the day this is a compliance industry, and we have to get those decisions right.

Jared S. Taylor
(00:09:26) You both mentioned something I want to touch on a little more. The technology is moving so fast. You talk about your early experiences and then how things continue to change. This is the question I always hear from founders and other leaders across healthcare. How do you keep up? Everyone has their own way of making sure they are not falling behind, and I know our audience would really appreciate anything you can share about how you stay ahead of it.

Deepak Ramaswamy
(00:10:02) Let me take that one. You are one hundred percent right, things change so quickly. But as founders, business leaders, and product owners who have a stake in building the product, we are always going to remain focused on the end problem we want to solve. If I have a problem of accurately extracting information from a document, that is the problem I want to be solving.

(00:10:35) From an architecture point of view, we design so that I can swap out this foundation model, or this fine tuned model, for something else. You define the solution in a way that lets you replace a piece later. When we first started the company we had an agreement with Hugging Face, the big open source AI model company, and we kept posing them the same problem. There are hundreds of thousands of these models out there. We are a small team. Which one should we pick? Which one should we take a bet on? We got a lot of guidance from them as the experts.

(00:11:20) But over time we figured out that we do not have to look at the shiny new object every time it shows up. We have to have some discipline and some cadence and remain focused on the end problems. Are we meeting the metrics and the thresholds? Then leave room in the architecture to swap open source weights for a foundation model, or the other way around. So we established a cadence internally, and the discipline not to look at the shiny new object. Of course there is a huge competitive advantage in evaluating what is new, so we go at it methodically and systematically, and we leave the architecture open so we can swap out one model, or a mixture of models, for something else.

Shaju Puthussery
(00:11:59) If I may add to that, Jared. We take that fundamental technical principle and look at our client base. What is the ROI for our client? What is the value? How can we take that same model Deepak was describing, the one we developed in 2022, and do it cheaper and faster?

(00:12:30) Credentialing is a good example, and you have a lot of experience operating in that space. We were able to do credentialing across all fifty states, get the license, take the screenshot from every state, in a matter of minutes. That took humans days. Log in, take a screen print, and so on. It evolved into agentic AI, but the point of doing it that way was to drive ROI for the client. Cheaper and faster. That is the fundamental innovation principle we use.

(00:13:00) With every model that comes out, whether it is agentic or otherwise, we ask what the value is, and we ask about the human experience. Simplify the process, do not complicate it. One touch, intuitive, with explainability. All of those guiding principles help us innovate faster.

Jared S. Taylor
(00:13:20) Are there areas where you two are critical? This can be a regulatory framework or something else. You have been able to do a lot, but what would you still like to see changed, let us say in the next five years?

Shaju Puthussery
(00:13:37) I would say change management is the biggest problem. It is not about technology and it is not about the solution. It is how we work with our target clients, whether payers, providers, or operators, and how quickly they can adapt and see the value. Quite often we spend a lot of time there rather than on the actual solution. Sometimes the solution is a big change and it is a shock for the end user, because they are used to doing a process one way.

(00:14:15) From a regulations point of view, we are NCQA accredited. Every systemic change we make to that credentialing process, going to those sixteen sources, whether it is DEA or NPDB, would take a long time if we went through the exact regulations each time. So instead we come up with processes that are quicker and faster and still compliant. We always go back to our NCQA leader in the company and say, we have implemented this faster acquisition of data, is this compliant? We will give you the screenshot, we will give you the date and time stamp, we will give you the authenticity of the source, which is what primary source verification requires. If the regulations could keep pace with the changing technology, that would be the ideal scenario, but regulations take much longer.

(00:15:10) That is the credentialing side. On insurance processing, we are able to auto adjudicate instantly. We adjudicate more than ninety percent of claims automatically without a human touch. But there are complex claims. Claims that require multiple sittings or multiple visits. Coordination of benefits, where a member has multiple benefits and you have to work out which is primary and which is secondary. And claims that require clinical review, which is where a company like Overjet comes in. All of those processes have to comply with regulations. There are clinical review regulations, insurance filing regulations, and prompt pay rules that require you to pay certain claims inside a certain window. So we still sit inside the perimeter of compliance, security, accessibility, and regulation, and we automate the process within it using agentic AI.

Jared S. Taylor
(00:16:04) As you were building out your model early on, was there a moment where the model was confidently, and maybe instructively, wrong in a way that taught you something real about the domain you are operating in?

Deepak Ramaswamy
(00:16:22) Maybe I can take a quick stab at it, from a business operating model point of view. When we first started the product, it was very focused on the insurance side, on how you configure benefits. A narrow sliver. We knew there were all these administrative problems going on in various companies, but we thought that one area, benefits configuration, was the most important.

(00:17:00) That still is important. But when you talk to clients, and we would demo and say, this is what we are thinking, this is the product, they would come back and say, wait a minute, I have this other problem. I am not able to credential providers in time. I am not able to load my files on time. I have a lot of duplicate records I cannot clean up.

(00:17:40) So what started at one end, where we thought from a business model point of view that this was the most important thing people were interested in, turned out differently. They are interested in everything. That informed the product strategy. We build so our customers can pick and choose the module that solves their need, instead of going with our initial hypothesis that there was one area they were interested in. That was obviously a daunting challenge, because as a startup you are making a big bet when you say you are going to do ten different things. But I think that has been successful for us, because it turns out customers are not just interested in one issue. They want to solve the whole problem.

Shaju Puthussery
(00:18:14) This is where the human comes into the loop. A lot of the time we can get those models to eighty five or ninety percent out of the box. How do you get from ninety to ninety eight or ninety nine percent, which is production ready? That is where we have a lot of humans looking at edge cases, repeating with more samples, building repeatability and reproducibility.

(00:18:45) So humans are a key component of the automation. I would say we can never do away with that. Humans and machines working together is where we will achieve the ultimate business goal.

Jared S. Taylor
(00:19:00) Sticking with you for a moment, Shaju. When I was doing my research, and correct me if I am wrong on this, you now run the credentialing rail for something like eighty seven percent of practicing dentists, which arguably makes you the holder of the most valuable provider data set in dental. So, putting on the credentialing hat, was credentialing always a Trojan horse to own the provider data layer, or is this something you backed into later? And part two, what responsibility comes with being the thing most of the system now quietly depends on?

Shaju Puthussery
(00:19:38) Great question. In fact, we accidentally stumbled into credentialing. We were working with one of the payers on a full blown claims adjudication system. I have been in the dental industry quite a long time and I thought credentialing was table stakes, something everyone could do. But then the CEO of the company said, Shaju, credentialing is a big problem. We are not able to onboard dentists in a timely manner. We do not have clean data. That is when Deepak and I spent a weekend looking at the problem together. Deepak called me on Monday and said, I have a solution for this. I am very honored to have a great tech partner who is always solving the problem ahead, once you know what the problem is.

(00:20:35) We run a platform, and to adjudicate a claim on that platform we need provider data. There are two hundred thousand active dentists in the US. Keeping that data clean, all the provider attributes from licenses to addresses to DEA, you name it, is a core part of our platform. We are able to monetize it through credentialing, which is an automated process, and we use the same information to adjudicate the claim.

(00:21:10) That is only one piece of the puzzle. We also have the patient data, the eligibility data, and myriad plan configurations, which are the benefits you and I have when we go to the dentist. What is my maximum, what is my deductible, all of that goes into it. Bringing all of those pieces together with AI as the central piece, cleansing and resolving issues, and now agentic AI, makes the platform truly remarkable. We are very fortunate that, as you said, we have more than eighty seven percent of US dentists on the platform, because we have very large payers who bring in all of that provider data and we are able to add value for them. We can get credentialing done in a few days at most, and within a day for the majority of providers, because of the automation and intelligence we have.

Jared S. Taylor
(00:22:06) A round of applause for you both and your team. That must be a really cool number to say inside the company. It is rare when you get to hear something like that.

(00:22:25) When I talk with people, it seems like healthcare administrative waste has floated in that same range for decades. As people playing your part in addressing that space, do you think the friction survives because someone actually benefits from it? Slow claims, denials, float on unpaid money. Who in the current system does not want the thing you are building, and how do you sell to them anyway?

Shaju Puthussery
(00:22:51) Naturally there is friction between the provider base and the payer. But AI and technology let us be a bridge between them. Previously there was more unwillingness on the part of providers working with payers, and payers working with providers. Now technology is enabling more automation. How do I get my prior authorization done quickly? How do I get my claims paid quickly? How do I adjudicate claims? As technology comes in, some of those natural barriers in the payer and provider space either coalesce or go away, and there is natural administrative automation happening on both sides.

(00:23:40) So I would say there are no winners here. We are not taking money from the payer and giving it to the provider. More automation means the dollars spent on administration get lower and lower. Can we take a portion of that money and put it toward patient care? That means shifting the focus of the dollar from admin to care outcomes. That is how we look at things. What is the best outcome we can produce for the patient?

Jared S. Taylor
(00:24:13) Interesting. Kicking it over to Deepak. I remember when I lived in Boston and would meet with VCs, it seemed like a lot of the health tech investors assumed dental was the beachhead and medical was the real prize, and medical claims are an order of magnitude messier than all that. Give me your honest read. Is dental a stepping stone to running these rails across all of healthcare, or is dental actually the better place to stay, and the rest of the industry, or at least some of the VCs I spoke with, are wrong about that?

Deepak Ramaswamy
(00:24:51) The answer to this depends on whom you ask. You will get a different answer each time. We have spoken to many, many VCs, obviously, doing fundraising, and you get a real diversity of opinion. Some VCs will not want to touch dental, maybe for the reason you mentioned. Other VCs, some of whom we were fortunate to work with, understand that this has been untapped. And if you expand the scope to what Shaju was saying at the beginning, we started off in dental and are going into vision and the ancillary benefits. You really can build an attractive solution in that space independent of medical.

(00:25:45) There are definitely a lot of synergies, but the key thing, especially for where we are as a company, is to have that focus. From a strategy point of view we always think about the point we need to get to, but this is where we are today. For us to be successful we have to demonstrate that we are providing good value in dental and all the ancillary spaces we talked about.

(00:26:20) And when we work on this, we are not thinking we just need to be good enough and then go on to the next thing. It is not, okay, we have demonstrated eighty percent of the value in dental, let us go on to medical. No. To what Shaju was saying before, there is a lot we can learn in dental and then apply to other domains. That being said, I have not been in dental nearly as long as Shaju, but I have worked in it for a number of years, and if I am being a bit generous, the technology is maybe several years behind medical. So sometimes the problems you solve here may not be directly translatable to medical, because the evolution is a bit behind.

(00:27:00) Even so, when we designed our system, we were not thinking dental only on day one. We were thinking about what happens if we get vision claims, and what happens if we get claims in these other ancillary industries. So the short answer is that it depends on whom you ask, and we were fortunate enough to center on investors who believed dental had potential and that the larger ancillary benefits industry had strong potential.

Jared S. Taylor
(00:27:20) If some of your VCs were on this interview with us right now, what would a few of their reasons be for backing you?

Shaju Puthussery
(00:27:29) I would say going deep into the same domain, solving the problem, demonstrating value through revenue, and having the capacity to make change in a conservative industry. How do you navigate compliance in healthcare, and security in healthcare? Some of the things Deepak and I did early on were getting SOC 2 Type 2 certified, getting HITRUST, and getting NCQA accreditation. We put the fundamental building blocks in place that allowed us to talk to large payers, because we had the foundations really well set. Then on top of that you bring the value added solution through AI and automation. So, go deep and innovate heavily.

(00:28:20) When I say innovate heavily, here is an example. We are working with a vendor serving around one hundred and fifty health plans, NationsBenefits. We are bringing fintech into the claims process, in real time. That means I can put your dental benefits on a flex card, just like your ID card. We all have a dental ID card when we go to the dentist. Now it also carries a Mastercard or Visa flex card loaded with your benefits. The dentist taps the card. Non covered procedures are not allowed, because the card knows what is covered under your plan and what is not. That is bringing fintech into claims adjudication.

(00:29:10) So we are going wider, and that is opening up business with health plans we could never get to before, and more business with vision plans. The center of it is expansion of innovation, and the core of it is AI and agentic AI that gives you cost reduction and faster time to market. Even though the flashier technology seems sexier, this takes much harder work to make it work. So that is how the VCs perceive us. Are you driving value and building a defensible business that lasts a longer period of time, rather than something flashy that gets quick revenue and then fails when the foundation gives way. We are very deliberate. It could be slower, but we are building on a different principle.

Jared S. Taylor
(00:29:55) I like it. One or two more things before we wrap up. We have hit on parts of this question, but it was one of the things I wanted to make sure I asked. As adjudication gets more autonomous, the accountability question gets sharper. I am not saying anything negative here, but this is something I hear from a lot of people. When an AI that is built wrongly denies or approves a claim, where does the responsibility actually sit? And can you also name one decision you deliberately refuse to let the model make, even though technically it could?

Shaju Puthussery
(00:30:34) I can take the first pass on that. Our models have to be transparent, auditable, and explainable. Each of those words is very important. Explainable means that whether it is the AI or a human, we should be able to explain exactly what step was taken and what reasoning was used to make a decision on a claim. Auditable means everything is written to a log file. We know at what point in time a particular claim came in, who touched it, and what algorithms were run. Transparent means it is not a black box that we are deploying.

(00:31:20) With all of those functions built into our automated models, the decision we do not allow is denying a claim. If a claim is denied for any clinical reason, there has to be a human in the loop to validate it. As Deepak said, a lot of adjudication is deterministic and rule based. I can have only two cleanings per year, or I have a two thousand dollar maximum on my benefit. Those are denials based on facts. But anything with a health outcome attached, we will not allow AI to make that decision. It has to be a human, a clinician.

Deepak Ramaswamy
(00:31:57) We draw a very clean boundary there. Where the clinician has to look at medical history or something like that, it is not our purview. That is our customers' decision, and they hire whomever they need to make it. But everything else that is deterministic and safe, that is not going to imperil a patient or produce a wrong outcome, and that is not going to lead to a HIPAA violation by sending the wrong information to the wrong patient, anything outside of that boundary is definitely within our purview and we strive to automate it.

Jared S. Taylor
(00:32:37) Very interesting. Thank you both. Last question before we wrap up. Let us put the future hats on for a second. Ten years out, what is the sentence or two you would need to be able to say to feel like you changed the system and not just the metrics? And is there a version of LightSpun that becomes the clearing layer for all of healthcare, not just dental?

Shaju Puthussery
(00:32:58) That is a fantastic question. We hope we innovate deeper in dental and automate things over the next five years to the level that an insurance company can run that layer without a lot of humans. We still need humans working alongside it. Beyond five years, can we translate this into health? Credentialing is an example. We do a lot of dental credentialing, and we are beginning to do medical credentialing, hospitals and facilities. So it is naturally expanding. We do payments using the flex card in dental. We can expand that to healthcare, the same card you tap when you go to your PCP. So our vision is to do really well in the lane we are in, go deep, establish good results and good value for our customers, and then begin to slowly expand into other healthcare domains.

Deepak Ramaswamy
(00:34:00) I would just add that it goes without saying that I would like us to be recognized as the company that automated the boring and the safe work away in dental and the ancillary industries, and yet recognized that humans are still responsible for the critical decision making. That has been the vision for us from day one. The boring stuff is where mistakes happen. Humans can make mistakes. If we can automate it safely, that has always been part of our mission.

Jared S. Taylor
(00:34:36) I really enjoyed our conversation here today, and I hope we can chat again in the near future.

Shaju Puthussery
(00:34:41) Jared, thanks again. This was a lot of fun for us.

Deepak Ramaswamy
(00:34:44) Jared, thank you for having us.