This Startup Is Automating America's Biggest Hospitals

In this episode of Founder Firesides, YC General Partner Aaron Epstein sat down with Kesava Kirupa Dinakaran, the Founder of Luminai (S20), which just raised a $38M Series B. Luminai is the AI transformation partner for health systems, automating the manual operational workflows for hospitals like Cleveland Clinic that still run on faxes and paper.

Dinakaran / YC — Luminai automating America’s biggest hospitals

YC General Partner Aaron Epstein interviews Kesava Kirupa Dinakaran (founder/CEO of Luminai, YC S20) on the YC Root Access channel — episode of Founder Firesides, published 9 April 2026, ~33 minutes. Frames a $38M Series B announcement and 4–5 years of company history; Luminai is the AI transformation partner for US health systems, building the agentic layer that converts unstructured inbound clinical/administrative paper-and-fax into structured EHR routing and scheduling for institutions like the Cleveland Clinic.

TL;DR

• Cleveland Clinic worked example as the canonical anchor: “the way that most patients get referred into the Cleveland Clinic is through a fax” — Luminai becomes the frontline inbox agent triaging incoming faxes (sales spam vs thank-you note vs high-criticality cancer patient), then extracts structured data, matches to patient/provider in the internal EHR, and routes to the correct one of “thousands of departments.” The wiki’s clearest founder-vantage operational anchor on the trillion-dollar US healthcare administrative-waste category.
• Horizontal-to-vertical specialisation as a credibility decision: Luminai began as a generalist horizontal automation platform; “north of 80% of our customers were in healthcare” in the data; Dinakaran made the big sort of decision to go all-in on healthcare, then narrowed further from “do anything for everybody in healthcare” to a very clear set of initial starting use cases and wedges. The verticalisation-as-trust-signal framing.
• Enterprise sales as a relational game, named as doctrine: “you’re not actually selling a customer like the Cleveland Clinic. You’re selling a champion within that one institution”; “really major institutions don’t really talk to you unless it’s someone who they trust is making the referral”. The wiki’s clearest founder-vantage articulation of the warm-intro discipline as structural-not-tactical at the hospital-C-suite altitude.
• The LinkedIn-contact-scraping script as the operationalisation: Dinakaran’s team wrote a script to crawl every investor / adviser / customer’s LinkedIn contacts and match against the target-customer profile; output 2–3 names per network; the asker then sends a tailored short blurb. Not cold email — relationship-graph traversal mediated by trusted-introducer-as-routing-hop.
• Founder-origin tangent that connects to the wiki’s existing canon: Dinakaran was the captain of the international Rubik’s cube team, broke world records, and on stage at the end of the interview solves a scrambled cube in 10.58 seconds. Casual mid-interview remark: “I learned Rubik’s cubes from Andrej Karpathy on YouTube, who eventually ended up founding OpenAI” — biographical bridge between the wiki’s existing Andrej Karpathy entity and the YC founder-vantage cluster.
• The reinvention thesis: “with every stage you have to reinvent yourself. Because the truth is, the next YC company is going to come in and pitch the most — the next dreamy vision that we were in five, four years ago.” Founder-vantage anchor on the continuous-pivot-at-scale doctrine for AI-native vendors against the next batch of YC competitors.

What was actually ingested

The full ~33-minute transcript (auto-generated English captions, ASR-cleaned). Cleanup applied to names (“Lumini” → “Luminai”, “Keshav”/“Keshab” → “Kesava”, “Peak 15” → “Peak XV”, “Erin” → “Aaron”, Rubik’s-related ASR variants normalised) and to punctuation/paragraph structure. Stage cues retained verbatim. The full raw transcript is at raw/videos/2026-04-09-dinakaran-yc-luminai-automating-americas-biggest-hospitals.md.

Substantive content

1. The Cleveland Clinic worked example (00:39–03:34)

The episode’s load-bearing anchor. The Cleveland Clinic does “a little over 16 million patient encounters” per year and treats some of the most complex care journeys; patients from around the world get referred in via fax from external physicians. Internal operational teams at Cleveland “figure out: is this sales spam… is this a thank-you note from some random provider, or is this a high-criticality cancer patient who needs immediate attention today?” — manual triage at scale.

Luminai is the frontline inbox agent: every fax hits Luminai first, becomes the initial triage layer; high-criticality goes straight through; non-urgent care gets “all the information [extracted], matching it to the right patient and provider within the internal EHR, and then routing it appropriately to the right department” (thousands of departments at Cleveland) before kicking off scheduling.

The thesis at the operational layer: “we’re essentially the data transformation layer for these institutions where we’re converting all of this unstructured data like faxes into structured data. And then we have a workflow engine on top where you can essentially build a set of verticalized agents to go solve very specialized and very important problems.”

The trillion-dollar framing: “30% of spend… on the administrative waste within healthcare which is over a trillion dollars.” Headroom claim, not measured impact — Luminai’s contribution against that pie is unaudited founder self-report.

2. The verticalisation decision (26:48–29:13)

Dinakaran narrates this as the company’s most consequential strategy moment. Luminai started as “a much more generalized and horizontal automation platform.” The realisation: “the moment you start to say ‘I can do anything,’ or ‘I work with someone else who doesn’t look like you at all,’ the credibility you have reduces pretty dramatically.” Data showed “north of 80% of our customers were in healthcare.” Two-step narrowing:

1. Vertical lock-in: all-in on healthcare.
2. Use-case lock-in within the vertical: not “do anything for everybody in healthcare” — but “a very clear set of initial starting use cases and wedges that makes them realize value incredibly quickly.”

The operational doctrine that follows: “these institutions are large and they don’t have time to think about what they can do creatively with your platform. You have to almost tell the story of what’s possible very very clearly, and you have to understand their problem so deeply that you’re able to talk to their problem like you’ve known them for 20 years.” The verticalisation-as-trust-signal framing is the founder-vantage instantiation of enterprise-ai-adoption — buyer credibility scales with apparent domain-specificity, not platform breadth.

The reinvention corollary (29:13–29:50): “with every stage you have to reinvent yourself. Because the truth is, the next YC company is going to come in and pitch the most — the next dreamy vision that we were in five, four years ago.” AI-native vendors face continuous pivot pressure from each subsequent YC batch.

3. Enterprise sales as a relational game (19:25–26:34)

The wiki’s clearest founder-vantage articulation of why warm-intro-discipline is structural, not tactical, at the hospital-C-suite altitude. Three interlocking claims:

(a) You sell a champion, not an institution. “You’re not actually selling a customer like the Cleveland Clinic. You’re selling a champion within that one institution who has to say ‘yes, I want to work with this product or team.‘” The strategy is to identify “forward-leaning people” — Dinakaran’s signal was “someone just growing rapidly through their career within that institution. There would be people sometimes in a hospital who’ve been there and they’re in their 30s but they’re like a C-suite executive at the hospital. That’s incredibly hard to pull off in that type of bureaucratic, like 100-plus-year institution.” Career-velocity as proxy for risk-appetite.

(b) Personal story is qualifying criterion #1. “when you lead with your personal story and when you lead with your history in life and you lead with your mission… even if you haven’t remotely accomplished it — they buy into that narrative if they believe it and they want to take a bet on you. That’s how these partnerships get done.” What the buyer is buying at founder-stage is the founder’s bet-worthiness, not the product. Companion to Aaron Epstein’s framing: “in the early days, there isn’t a company or a brand or social proof of all these other customers using you… they have to take a bet on you.” Dinakaran: “That’s what they’re buying — not even your product — is that you’re going to make it right.”

(c) The LinkedIn-contact-scraping script as the operational machine. “we wrote a little script where we would go through everyone’s LinkedIn and scrape every single contact of theirs, and then just match it against our target customer profile. And out would come maybe two or three names. And then those two or three names… [send] a short blurb.” The asker did not need executives — “if it was a target customer we really wanted to get in front of, I was like, ‘If I can just find someone at the company, I’m sure that we can pitch it to them and get them excited that then they might be willing to get us maybe two layers above and then slowly make our way to the actual person who’s the champion.‘” Relationship-graph traversal mediated by trusted-introducer-as-routing-hop — explicitly contrasted with cold outbound.

The cadence: “if you do that iteratively over many many many many weeks and months, you’ll have a very very thick pipeline of people who are willing to have deep conversations with you.” And: “the world is much smaller — especially if you focus your customer base — than I expected it to be.” The smallness-of-the-world observation as the operational reward of vertical focus.

The red-eye discipline: “the moment someone was willing to meet with me, didn’t matter where they were in the country, I’d be like, ‘I’m happy I’ll show up at, you know, wherever you are tomorrow morning.’ And the number of red-eyes I’ve taken to make that type of thing happen, or the number of red-eyes our team has taken… is just more than we can count.”

4. The pivot template (16:39–19:25)

Dinakaran’s contrarian view: pivoting can be methodical, not artful. “At the end of the day, when you’re building something for a business customer… there’s actually a lot more science than art.” Two threads in parallel at the beginning of the pivot search:

1. Mission/story alignment: “is there a mission or story that truly energizes me to work on a problem for a very long time?” A serious family health event was the trigger for healthcare.
2. Customer-pain saturation: “as we started to talk to larger and larger enterprise organizations, this whole problem of operational teams mostly doing deeply manual workflows became very evident.” Generalise across enough conversations to pull the underlying problem signal.

The original Luminai application to YC was for developer documentation automation — “building documentation for engineers and making it easier for engineers to create documentation” — which Aaron Epstein flagged at the original YC interview as “not a great idea.” Three weeks of customer iteration confirmed it. The pivot-to-healthcare-ops emerged from the two-thread search above.

5. Founder-origin: Rubik’s cube, the Karpathy-as-tutor tangent, and the United World College fellowship

Dinakaran spent his childhood (~age 11 onward) ~7 hours/day on Rubik’s cubes; broke world records; was the captain of the international Rubik’s cube team. Two motivations he names retrospectively:

1. Level playing field: Rubik’s cube community was a meritocracy where “you’re 11 years old and you hadn’t done well academically or something” didn’t matter; doctors, CEOs, musicians, and 65-year-olds played by the same rule (“can you solve a Rubik’s cube fast?”).
2. Tight measurable feedback loop: “every day, the seven hours I practiced, it was so clear what my feedback was — I landed at this time, why did it take this much time, what do I need to do to get better.” The dopamine of fast feedback iteration loops — a structural echo of the agentic-engineering eval-as-gradient framing the wiki already carries.

The tangent that bridges to the wiki’s existing canon: “I learned Rubik’s cubes from Andrej Karpathy on YouTube, who eventually ended up founding OpenAI.” Casual aside, but the biographical bridge from Dinakaran’s Rubik’s-cube-as-childhood-skill to Karpathy’s later AI-frontier work is a non-trivial trace through the wiki’s existing entity graph.

The cycle from Turkey to China (high-school fellowship project at the United World College) was the credential that flew Dinakaran to Silicon Valley for a 10-day trip on a tourist visa — and he never left. The hackathon-arbitrage period (winning $40-45K across 5-6 hackathons by iterating with the organisers as the de facto judges in real time — “by the end of the hackathon, in 48 hours, you’ve gotten like maybe whatever — 16 iteration loops with them, back to back to back to back”) is a separate but consonant fast-feedback-loop anchor.

The closing aphorism (30:01): “The level of aggression and ambition you can have is probably a hundred to a thousand times larger than you think you should have. And it has nothing to do with your experience or where you come from. There are no rules — outside of integrity rules — with the types of problems you can go out and solve.”

6. Operational scale and capital signals

• Lifetime capital: ~$60M raised (confirmed by Aaron Epstein at 32:00).
• Latest round: $38M Series B, led by Peak XV (formerly Sequoia Capital India), 2026.
• Customer pattern: large academic medical centres including the Cleveland Clinic. 80%+ of customers in healthcare.
• Cleveland Clinic per-year scale: ~16M patient encounters.
• Industry context: US healthcare administrative spend = ~30% of total US healthcare spend = on the order of $1T (Dinakaran’s framing).

7. Cross-cutting observations

1. Paper-and-fax-vertical-AI-wedge as a named pattern. Paired with Vori (grocery), the wiki now carries two trillion-dollar-industry-scale exemplars of the paper-and-fax → AI-native operational layer template. Same rhetorical move: name the analog substrate, describe the wedge, explain why the industry stayed analog.
2. Verticalisation-as-trust at the enterprise altitude. Dinakaran’s “the moment you start to say ‘I can do anything’… the credibility you have reduces” is the founder-vantage operational complement to Hu / Stanford CS153’s prescriptive “closed-loop company” doctrine — verticalisation is the buyer-side trust mechanism, closed-loop is the internal-operations mechanism.
3. Founder-led sales survives the AI era at the enterprise altitude. The enterprise worked-example anchor of the wiki’s founder-led-sales concept. Companion to Campfire’s “I do really recommend founders to stay in the founder sales mode… offloading it to AI, offloading it to some AE, [feels like] let’s just bring in a professional” — Dinakaran’s LinkedIn-contact-scraping script + warm-intro discipline + red-eye cadence is the operational how of staying in founder sales mode in the AI era at the hospital-C-suite altitude.
4. The Karpathy biographical bridge. Casual mid-interview tangent connects Dinakaran’s Rubik’s-cube origin to Andrej Karpathy’s YouTube tutorials. Not load-bearing analytically, but a small graph-traversal-pleasure connection — the wiki’s existing entity is named directly by a YC founder fireside published five years after the original tutorials.
5. Hackathon-arbitrage iteration loops as a founder-development primitive. The 16-iteration-loops-in-48-hours pattern Dinakaran ran with hackathon organisers as de-facto judges is a small but clean fast-tight-feedback-loop anchor — a behavioural pattern (talk to the judge every 2 hours, rebuild based on feedback) that recurs at startup-pivoting scale and at agent-engineering scale (eval-as-gradient).

Linked entities and concepts

Entities directly named or substantively discussed in the source:

• Y Combinator — accelerator; Luminai is S20. Aaron Epstein is the interviewing GP. Source-count bumps 7→8.
• Aaron Epstein — Dangling first mention (this is his first-source appearance; the YC General Partner conducting the interview).
• Kesava Kirupa Dinakaran — Dangling first mention (Luminai founder/CEO).
• Luminai — Dangling first mention (the company; YC S20; healthcare-administrative AI).
• Cleveland Clinic — Dangling first mention (the load-bearing customer worked example).
• Peak XV — Dangling first mention (Series B lead investor; formerly Sequoia Capital India).
• United World College — Dangling first mention (high-school fellowship that funded the Turkey-to-China cycling trip).
• Andrej Karpathy — substantive cross-reference (Rubik’s-cube YouTube tutor → eventual OpenAI co-founder). Existing entity. Source-count: increment.

Concepts touched substantively:

• enterprise-ai-adoption — the trillion-dollar US healthcare administrative-waste anchor and the founder-vantage verticalisation-as-trust mechanism. Source-count: +1.
• agent-harness — the workflow engine on top + verticalized agents phrasing is a vendor-product instantiation of the harness concept. Source-count: +1.
• automation-vs-augmentation — the people, process, and paper → structured agentic workflow framing is automation at the operational layer. Source-count: +1.
• ai-employment-effects — operational teams whose “entire job is basically to look at these faxes” are the population whose work Luminai replaces or transforms. Source-count: +1.

Dangling (single-source mentions, deferred): Aaron Epstein, Kesava Kirupa Dinakaran, Luminai, Cleveland Clinic, Peak XV, United World College.

Caveats

• YC Root Access Founder Fireside with obvious commercial motive (Series B announcement framing); all financial and impact claims are founder self-reports, unaudited.
• “30% of admin spend on administrative waste” and “over a trillion dollars” are TAM-framing figures cited by Dinakaran without methodology — directional, not bookable.
• The “16 million patient encounters” annual figure for the Cleveland Clinic is a checkable public statistic; not independently verified at ingest time.
• Customer logos beyond the Cleveland Clinic are not named on-stage; the lifetime customer breadth is inferred from “80%+ healthcare” and “a dozen workflows across the Cleveland Clinic and other health systems.”
• The Series B = $38M / lifetime = $60M breakdown is paraphrased from on-stage discussion; not cross-referenced to public funding databases at ingest time.
• The Andrej Karpathy / Rubik’s-cube tutor remark is a single founder-stage line and not load-bearing; treat as colour, not evidence of any institutional connection.
• Cold-outbound rejection is a founder doctrinal claim — multiple Luminai customers may still arrive that way; “probably some major portion of our customers still come through that way” (Dinakaran) leaves the door open even within the same answer.