What if rules gained context and reasoning? What if a machine earned the right to act? What if the policy fit the person? What if care continued between visits? What if the $1T spent on admin went to Care.

Kā is the system of record for healthcare's decisions. Frontier models supply reasoning. Kā compounds domain judgment and converts it into evidence-backed, technically enforced, revocable authority to act. Proven decision families become standalone companies.

Frontier capability
Selectively rented or hosted
Domain judgment
Owned in rules, adjudicated exceptions, evaluations, calibrators, verifiers, and narrow models
Decision authority
Earned through evidence, enforced at runtime, and revoked when performance deteriorates
01 · Mission

Return $1 trillion from admin and waste to Care.

Coverage decisions, coding, compliance, adjudication, and appeals consume resources that never reach a patient. Kā builds the shared decisioning layer beneath every workflow that can return them. Every decision it automates, audits, and improves is one that used to consume a clinician's time, a member's attention, or a dollar that never became Care.

Returning that trillion takes more than cheaper software. It takes healthcare that can reason in real time.

The next decade of healthcare will be built on systems that reason in real time.

P01 · Architecture
add-onengine

AI moves from a copilot beside the workflow to the decision system underneath it. The layer the operator runs on.

P02 · Logic
rulesrules + reasoning

Rules set the boundary. Reasoning resolves the context. Compiled rules define the permissible decision space; reasoning adapts to evidence and state within it.

P03 · Specificity
policyperson

Population evidence guides care. Individual state determines how it should apply now. AI-native care fits the decision to this person, this episode, this moment.

P04 · Temporality
episodiccontinuous

Care today happens at visits, claims, and reviews. AI-native care thinks in the background, between encounters. Closer to monitoring than to documentation.

Real-time reasoning has to live somewhere it can be governed. Kā is that layer: neutral by design, built to operate across carriers and captured by none.

Healthcare's systems of record hold the data. Kā is the system of record for the decisions.

A governed layer that turns policies, contracts, clinical context, and longitudinal data into decisions that can be computed, audited, challenged, and improved. Every decision emits a trace: inputs gathered, rule and authority in force, exception granted, human who approved, downstream outcome. Stitched across members, providers, contracts, and time, those traces form a context graph. Precedent turns searchable. Calibrated abstention becomes a first-class primitive. Every decision carries the reason it was permitted. Frontier capability is rented and replaceable. Domain judgment compounds in owned artifacts. Authority is earned, enforced, and revocable.

03.1 · Architecture

A shared decisioning layer carries the context that every workflow today rebuilds from scratch.

From fragmented stack to a unified decisioning substrate.

Today's systems are stacked vertically and assemble decisions out of disconnected pieces. The substrate sits beneath them: a single decisioning layer under every workflow, claim, and care plan. Existing tools remain in place as interfaces over it; context stops being rebuilt from scratch.

Today
Isolated decision stacks
EHR + rules Prior authorization
Claims + policies Adjudication
Notes + care plans Longitudinal care

Each workflow rebuilds its own context.

With Kā
Compounding integrated judgment
Shared sources
EHR
Claims
Policies
Clinical evidence
Patient history
Decisioning substrate
Context graph
Authority + evidence
Abstain + escalate
Trace + learning
Decision families
Utilization management
Care management
Network management
Adjudication
Pricing
Underwriting
Personalized Dynamic Real-time

One governed context. Many controlled workflows.

03.2 · Capability model

What if AI had to earn the right to decide?

The autonomy ladder: ceiling is law, rung is evidence.

Every decision family gets an autonomy ladder: five rungs of machine authority under a legal ceiling. The ceiling is what statute, regulation, contract, or tenant policy permits. The rung is what the system has earned through blind shadow, calibration, and outcome evidence. One family may be capable of Rung 4 and capped at Rung 2 by law; another, permitted to Rung 3, may still sit at Rung 1 for lack of evidence. Every promotion is signed by a certificate; every demotion is automatic.

R0
Blind shadow Kā · Today
The system observes and produces decisions; output is withheld until the human decision locks. Both lock before comparison, so anchoring is avoided by design.
Machine sees, human decides
R1
Assisted review
Model output is visible; the human decides. Certified separately from blind shadow, because the review posture is different.
Machine drafts, human decides
R2
Draft decision
The system proposes a complete decision with full rationale. The human approves before execution. First cross-family promotion target for guideline-concordant volume.
Machine decides, human approves
R3
Limited auto-execution
The system executes low-risk validated actions inside explicit guardrails. Escalation on any uncertainty or harm-class exposure threshold. Compiled rule artifacts become mandatory at this rung.
Machine executes, human audits exceptions
R4
Auto-execution + post-hoc audit
Full auto-execution within the domain. Human review is post-hoc and audit-driven. Requires certified performance across two or more tenant distributions. Cross-carrier generalization becomes an engineering prerequisite.
Machine operates, human sets policy
C
Ceiling Prohibited
Legally, contractually, or clinically capped. No evidence permits the system to operate above it. Adverse medical-necessity determinations, for example, hold at licensed-clinician authority regardless of ladder evidence.
Human only, by law
03.3 · Build system

Authority is earned: evidence promotes a decision family a rung; failure demotes it automatically.

The certificate loop: how a decision family earns its authority.

The autonomy ladder defines what a system is allowed to do. The certificate loop defines how it gets there. A decision family enters in blind shadow. Every event emits a trace. Human and machine decisions are compared after both lock. Random audit samples are adjudicated into gold sets and certify performance; disagreement cases are adjudicated separately and drive learning. Calibration is measured against observed error. When evidence clears the family's promotion gates, an autonomy certificate grants a rung of live authority. Below the legal ceiling. Scoped to a tenant. Evidence promotes. Failure demotes, automatically.

Stage 01
Exploration

We test whether a decision family can be made computational at all. Blind shadow only; no live authority granted.

Is there enough signal, structure, and consequence?

Stage 02
Constrained deployment

The system runs inside bounded environments under a first certificate. Human-in-the-loop. Full traceability. Measured against gold sets, adversarial sets, and outcome-linked evidence. Calibrated abstention on every uncertainty crossing.

Does it hold rung under drift, under adversarial pressure, under audit?

Stage 03
Autonomous operation

Recertification runs automatically as models, rules, and authorities change. New model generations translate into expanded authority without re-architecture. Certification lag becomes the differentiation metric.

Does it earn the right to operate, and can it be recertified faster than the alternatives?

If successive deployments do not make the next safe delegation faster, cheaper, and more trustworthy, the thesis is not working.

We build systems that earn the right to operate.

One substrate holds many decision families. The Studio turns each proven one into a company.

Every category-defining company starts as a question.

01 · Ask
"What if…"

Each company begins as a question about a decision healthcare gets wrong at scale. The same question the substrate is built to answer.

02 · Refine
Sharpen the question

We narrow it to a single decision family and fix its legal ceiling.

03 · Prove
Earn the authority

Blind shadow, calibration, and outcome evidence establish that the substrate can hold the decision before any live authority is granted.

04 · Launch
Spin out a company

A proven decision family becomes its own company, operating on the substrate from its first day in market.

Most questions do not survive the funnel. The ones that do compound: each company sharpens the substrate the next one is built on.

We start where decisions are dense enough to evaluate, consequential enough to matter, and bounded by a clear legal ceiling.

The substrate proves itself inside specific decision families. Each has its own rules, its own regulatory ceiling, and its own path up the autonomy ladder.

T01
Computable contracting

Provider contracts compiled from static fee schedules into executable, parameterized functions. Episode-level rates form at prior authorization from clinical context, provider quality, and capacity. White-box by construction.

T02
Member continuity

A continuity graph maintained year-round for every Medicare Advantage member: providers, prescriptions, network access, out-of-pocket exposure. It surfaces plan disruption before it happens.

T03
Longitudinal care orchestration

Decisions that require coherent state across encounters, providers, and years. State-snapshot contracts and reproducible replay make the substrate the closed-loop instrument on the trajectory.

T04
High-impact intervention routing

Calibrated stratification of the small share of members or episodes where a specific action materially changes the outcome. Conformal risk bounds per stratum. Verifier-panel routing on the disagreement stream.

Founders who have built and exited in health tech, run risk-bearing care at scale, and put reasoning models into regulated production.

A founding cohort across frontier ML, clinical operations, and risk-bearing economics.

Puneet Maheshwari
Founder · CEO
LinkedIn ↗

Chief architect of the substrate. Founded DocASAP and scaled it into a category leader in patient access before its acquisition by Optum, where he went on to build Optum Real. Earlier, McKinsey. Wharton MBA and a computer-science foundation. Leads the certificate loop and the studio's technical doctrine.

Brady Priest
President and COO
LinkedIn ↗

Leads partnerships, portfolio, and the operating cadence of the studio. Prior operating leadership at Clover Health and across UnitedHealth Group, spanning payer strategy and risk-bearing care operations. He has scaled regulated healthcare organizations through their sensitive early years.

Five open problems, and the roles that will solve them.

The questions a frontier operating layer for healthcare has to answer.

None of these has a textbook solution. We are hiring the people who want to do the work.

Q01
How does a reasoning system warrant autonomy in a regulated, life-critical domain?
Safety · evals
Q02
What does interpretability look like when the audit unit is a clinical trajectory rather than a single inference?
Interpretability
Q03
How do you compile a payment contract into an executable function without losing legal correctness?
Programming languages
Q04
How do agents maintain coherent state across years of patient context, across handoffs and gaps?
Memory · long horizon
Q05
What evaluation infrastructure delivers RCT-grade evidence at the speed of software?
Causal inference

We are hiring engineers, researchers, clinicians, and operators building the AI-native decisioning layer of healthcare. Write directly.

puneet@kalabs.io