CHAPTERS
- 0:05 – 0:57
$55M Series B to build an agentic AI platform for health systems
Nish Khandwala announces Bunker Hill Health’s $55M Series B led by Khosla Ventures with participation from YC and others. He frames the fundraise around accelerating adoption of “agentic AI” in hospitals via a platform approach rather than one-off tools.
- •$55M Series B announcement and key investors
- •Capital will fund an agentic AI platform for health systems
- •Thesis: platform-based adoption is the future of hospital AI
- 0:57 – 3:15
The core problem: hospitals can’t iterate on AI fast enough
Bunker Hill Health focuses on lowering the cost and increasing the speed of iteration in healthcare. Nish contrasts clear AI potential with the reality that hospitals struggle to capture it in practice.
- •AI value is obvious to hospital leaders and clinicians
- •The real bottleneck is operationalizing AI, not proving potential
- •Goal: make experimentation cheap and turning off failures easy
- 3:15 – 5:46
Why onboarding a single AI tool can take ~2 years
Nish breaks down the practical barriers that stretch deployments into year-plus timelines: procurement, security, privacy, IT bandwidth, and change management. These frictions make adopting many point solutions unrealistic even when teams want them.
- •Procurement, contracting, security, and privacy reviews are slow
- •Shared IT resources create long queues and delays
- •Change management and rollout complexity extends timelines
- •Pain outweighs benefit when each tool solves only one narrow task
- 5:46 – 10:12
Innovator burnout inside hospitals: great ideas that go nowhere
From a clinician/innovator perspective, LLM-based ideas (e.g., finding eligible MASH patients for new therapies) are straightforward but repeatedly blocked by organizational constraints. The result is a demoralizing cycle where innovation stalls and enthusiasm fades.
- •Clinicians see clear use cases (e.g., eligibility outreach for Wegovy in MASH)
- •Pushback is often structural (bandwidth/process), not anti-innovation
- •Repeated failure to deploy tools creates burnout
- •Lower iteration cost unlocks more clinical and operational improvements
- 10:12 – 10:51
Bunker Hill’s platform approach: connectors + AI + automated actions
Nish explains how Bunker Hill positions itself as a platform that both reads from and writes to hospital systems. The goal is to avoid dashboards-only outcomes and instead drive real operational follow-through via automated actions.
- •Platform integrates with EHRs, ERPs, imaging, payer/registry portals, etc.
- •Supports LLMs and specialized models (e.g., vision/diagnostics)
- •Emphasis on closing the loop: outputs must trigger actions, not just reporting
- 10:51 – 13:01
The three pillars: Knowledge, Reasoning, Action (and “Carebricks”)
Bunker Hill formalizes the platform as three building blocks: Knowledge (ingest), Reasoning (AI), and Action (execution). Examples of “action” include patient outreach and provider messaging through existing hospital channels—even fax.
- •Knowledge = connectors to systems of record + internet guidelines
- •Reasoning = AI layer that produces decisions/insights
- •Action = automated execution (mail, SMS, voice, EHR messages, portals/fax)
- •Designed to be buildable/self-serve over time, not bespoke per use case
- 13:01 – 17:31
Origin story: from Stanford AI research to real clinical deployment pain
Nish recounts starting in AI research (semi/unsupervised learning) where healthcare was initially “just a dataset.” A preventive cardiology collaboration revealed a powerful use case—detecting coronary artery disease from incidental CT findings—yet deployment proved surprisingly hard.
- •Background: AI efficiency research; healthcare initially a compelling dataset
- •Preventive cardiology use case: identify plaque on existing CTs to prevent heart attacks
- •Model building felt feasible; deployment was the real obstacle
- •“It’s someone else’s job” highlighted the institutional gap
- 17:31 – 18:45
Personal catalyst: his father’s heart attack makes the mission urgent
After years of struggling to get a preventive tool deployed, Nish’s father experiences a heart attack with the same type of CT-based findings the tool targeted. The personal and professional frustrations combine, pushing Nish and his cofounder to focus on systemic deployment barriers rather than a single algorithm company.
- •Father’s heart attack mirrored the ‘missed plaque on CT’ problem
- •Reinforced that care gaps are common and costly
- •Cofounder faced similar deployment dead-ends (robotic surgery AI)
- •Decision: solve deployment/iteration as the broader platform problem
- 18:45 – 22:05
LLMs change the game: many ‘algorithms become prompts’ and broaden impact
LLMs both intensify the demand for deployment (more ideas become easy to build) and simplify the infrastructure needed to support many use cases. Nish describes a “Maslow’s hierarchy” of hospital problems and argues LLMs let platforms address everything from basic ops to “enlightened” clinical applications.
- •LLMs increase the number of viable internal hospital use cases
- •Shift from bespoke ML pipelines to prompt/tool-based workflows
- •Platform can serve ‘base of Maslow’ operational needs and higher-order clinical goals
- •Solving boring foundational work enables adoption of cooler innovations
- 22:05 – 23:35
First customer by cold email: Cleveland Clinic and the birth of ‘Action’
Bunker Hill’s first major deployment came via a cold email to a Cleveland Clinic physician. Scaling insights revealed that simply flagging patients isn’t enough; hospitals need automation to manage follow-up at volume, which drove the platform’s action layer.
- •Cold outreach landed a flagship early adopter (Cleveland Clinic)
- •Retrospective studies surfaced thousands of actionable patients
- •Delivering a list isn’t workable—follow-up staffing is a constraint
- •Action layer emerged to automate outreach and care navigation
- 23:35 – 26:01
Finding the right abstraction: one paradigm across clinical and admin workflows
Nish explains how knowledge→reasoning→action maps across diverse hospital problems: care gaps, insurance eligibility, revenue cycle coding, prior auth, and more. This shared structure enables rapid expansion to many use cases without rebuilding from scratch each time.
- •Same abstraction applies to Medicaid eligibility, billing/coding, care gap closure
- •Supports both clinical and operational/financial workflows
- •Key insight: reusable primitives beat bespoke implementations
- •Explains why they can tackle 15–20 use cases per health system
- 26:01 – 27:46
Engineering for speed: primitives that non-engineers can deploy
The hardest technical work is choosing the right primitives so forward-deployed teams can configure solutions without sending requests back to core engineering. This design makes iteration fast and prevents the platform from devolving into slow custom services.
- •Goal: reduce iteration time by avoiding engineering-in-the-loop per use case
- •Forward-deployed teams use the platform directly to build/adjust workflows
- •Core engineering focuses on platform abstractions, not customer-specific code
- •Enables unusually high number of deployed use cases per customer
- 27:46 – 34:04
Hospitals are technologically similar but operationally different
Nish argues that systems and standards (HL7/FHIR, imaging formats) are relatively consistent across large hospitals, but workflows vary widely. The platform must adapt to local processes rather than force change management-heavy standardization.
- •Standards like HL7/FHIR are widely adopted; imaging is standardized
- •Big differences show up in processes and clinical preferences
- •Example: lung nodule pathways differ (Fleischner guidelines vs Brock score)
- •Platform must ‘rearrange the bricks’ to match each hospital’s workflow
- 34:04 – 39:27
Reliability: tool use, citations, and benchmarking vs standard of care
They address hallucinations through tool use (dynamic lookup of guidelines/policies) and by requiring citations back to systems of record to make outputs auditable. Nish emphasizes evaluation against the ‘standard of care’—often a lower and more realistic bar than perfection, especially when the alternative is doing nothing.
- •Tool use reduces hallucinations by letting models fetch up-to-date sources
- •Architectural guardrail: no claims without traceable citations to the EHR
- •Example: registry question answered via a niche note reference (vaping)
- •Performance should be compared to standard of care, not 100% accuracy
- 39:27 – 43:30
Building a lean team (21 people) around platform primitives and deployment
Nish describes a team design where engineers work only at the platform layer and forward-deployed staff drive adoption across many stakeholders and use cases. AI and coding agents reinforce a preference for smaller, more product-oriented teams, with planned growth to ~35 rather than hyper-scaling headcount.
- •Engineers focus on abstractions/primitives, not individual hospital use cases
- •Forward-deployed team maps to hospital stakeholders (IT, procurement, users, execs)
- •Small teams reduce coordination overhead and keep customer context close
- •Current scale: 21 people; target ~35 by year-end
- 43:30 – 47:50
Patient-facing future: the ‘turkey bone’ journey and an always-on care system
Nish walks through a real patient story where an incidental finding on a CT triggers automated outreach and accelerated cardiology care. He extends the vision to end-to-end agents supporting diagnosis, referrals, scheduling, prior auth, follow-up, and billing—creating a more connected experience where someone is always ‘looking out’ for the patient.
- •Incidental plaque detection triggers automated patient notification
- •Agents reprioritize referral queues by urgency and care modality (virtual vs in-person)
- •Agents assist scheduling and payer approval (prior auth/coverage)
- •Outcome: earlier intervention (triple bypass) and better continuity of care
- •Long-term: better trial matching, earlier diagnosis, and personalized therapy selection
