Shaju Puthussery and Deepak Ramaswamy are the CEO and CTO of LightSpun, which is rebuilding the back-end engine of insurance processing as agentic AI infrastructure, starting in dental and moving into vision and ancillary benefits. Both came from Overjet, where Deepak was a co-founder and Shaju the first employee, and this conversation is largely about why they walked away from that thesis. Reading X-rays was the visible AI problem. Underneath it was a plumbing problem: claims that never reach clinical review because the documents were wrong, the provider record did not match, or the file never loaded cleanly.
Two things in this episode surprised us. The first is that their most valuable asset was an accident. Shaju assumed credentialing was table stakes until a payer CEO told him it was blocking dentist onboarding, and one weekend later Deepak had an approach. That became the rail for roughly 87% of practicing dentists in the US and the provider data spine that feeds adjudication. The second is Shaju's answer to whether anyone profits from claim friction, which is not the cynical answer most people give.
We discuss:
Why Deepak argues the AI question is not either-or, and why world-class clinical review is worthless if the claim cannot get to it
The moment their business model was confidently wrong: they built for benefits configuration, customers came back asking about credentialing, duplicate records, and file loads, and a startup that planned to do one thing had to bet on ten
Why credentialing was never a Trojan horse for the provider data layer, how the same 200,000-dentist dataset gets monetized twice, and what obligation comes with being the thing the system quietly depends on
The honest ceiling on model performance: 85 to 90% out of the box, and why the climb to 98 or 99% production-ready is where the humans actually live
The exact decision they will not automate, with the line drawn between deterministic denials (two cleanings a year, a $2,000 annual max) and anything touching a clinical outcome
Why regulation, not technology, sets the pace, and how they take a faster primary source verification method to their internal NCQA leader with screenshots, timestamps, and source authenticity to prove it still holds up
Shaju's contrarian read on the $17 to $21 billion admin waste question: no one is winning from the friction, both sides are automating, and the real goal is shifting dollars from admin to care
Bringing fintech into adjudication with a benefits flex card that carries a Visa or Mastercard rail, blocks non-covered procedures at the chair, and opened doors to a vendor network of roughly 150 health plans
The discipline of not chasing every model release, what Hugging Face taught them early about picking bets, and why the architecture is built to swap foundation models out entirely
Deepak's pushback on the beachhead narrative: dental is several years behind medical, which means the solutions may not transfer cleanly, and they designed for vision and ancillary from day one rather than treating dental as a waypoint
Why compliance came before the AI story, with SOC 2 Type 2, HITRUST, and NCQA in place first so they could get in the room with large payers at all
The legacy Deepak actually wants: recognized as the company that automated the boring and the safe, and left the critical decisions with people
—
—
Brought to you by: Sage Growth Partners — Value-focused strategy and marketing for growth-driven healthcare organizations.
—
