by Malay Baral, Founder and CTO, AdiOS Platform
"Sovereignty is necessary. It is not sufficient."
The sovereign AI conversation has matured fast. In two years it has moved from fringe concern to global infrastructure priority. Nations are building sovereign GPU clouds. Enterprises are demanding data residency. Regulators are encoding these requirements into law.
This is progress. It is also incomplete.
Data residency answers one question: where does the data live? There is a second question that is harder and more important: what happens to the intelligence after the data stays inside the walls?
AdiOS and the COS framework are built around one answer to that question.
The Gap That Residency Does Not Fill
When we talk about sovereign AI today, we are mostly talking about infrastructure. Compute that stays inside national or institutional boundaries. Models that run on premises. Data that never crosses a jurisdictional line.
These are real and necessary properties. But they describe a constraint on data movement, not on intelligence ownership.
Consider a typical enterprise AI deployment. The operator's data stays inside. The inference happens inside. The answer comes back inside. But the intelligence that produced that answer lives in the vendor's platform. Every time the operator runs a query, they are renting intelligence they do not own.
When the contract ends, the operator keeps the data. The intelligence resets.
These are not theoretical gaps. They are operational realities that every regulated Indian enterprise will face as DPDP Act enforcement arrives.
Figure 1: What data residency answers, and what it does not
Introducing the COS Framework
COS stands for Circular Operating System. It is a practitioner framework for designing and deploying AI infrastructure where intelligence compounds inside the institution that generates it.
COS is not a product. It is the blueprint. AdiOS is the operating system built on it.
The framework rests on a single axiom:
Every operation is a circular transaction where value circulates within the community that generates it.
This axiom produces five consequences that the framework enforces:
Figure 2: The Circular Axiom and its five structural consequences
Each consequence is a design constraint. An AI deployment that violates any of these five properties is, by COS definition, extractive rather than circular.
How COS Defines Circular Intelligence
COS organises AI deployment around nine knowledge areas that a practitioner must address. These span sovereign data governance, intelligence architecture, distributed systems, regulatory compliance, last-mile deployment, AI operations, knowledge engineering, ecosystem participation, and organisational readiness.
The framework provides three assessment deliverables for any institution:
Figure 3: COS Engagement Model
The assessment framework uses seven hypotheses to evaluate an institution's current AI posture:
- H-1 Sovereignty Posture. Does data stay within the defined boundary?
- H-2 Extraction Risk. How much knowledge is leaking to external vendors?
- H-3 Offline Readiness. Can the institution operate with zero network?
- H-4 Compliance Architecture. Is compliance structural or procedural?
- H-5 Knowledge Compounding. Does knowledge survive staff turnover?
- H-6 Ecosystem Participation. Is the institution a consumer or a contributor?
- H-7 Last-Mile Coverage. Does AI reach the operational edge?
Most institutions today score well on H-1 and poorly on H-3 through H-7. The COS gap analysis makes this visible and actionable.
The Circular Loop in Practice
The core operational pattern in COS is the Circular Learning Loop. It is the mechanism by which every operation contributes to institutional intelligence rather than simply producing an answer and discarding the context.
Figure 4: The Circular Learning Loop
This loop runs at three tiers of confidence. Personal observations at the operational edge. Domain knowledge validated across a team or facility. Enterprise intelligence proven across the institution and ready to compound across the network.
Raw operational data never leaves the sovereign boundary. What circulates through the network is scored, validated intelligence, gated by consent.
What Circular Intelligence Means for Indian Enterprises
India presents a combination of conditions that does not exist anywhere else simultaneously.
DPDP Act 2023 full enforcement arrives in 13 months. The penalty structure runs from INR 50 crore to a maximum of INR 250 crore per breach, depending on violation type. This is not a future risk. It is a present architecture obligation.
The scale of the regulated market is significant:
- 12 PSU banks. 21 private sector banks. Approximately 9,300 NBFCs.
- Around 740 million ABHA IDs. 1.5 million health facilities.
- Between 100 and 150 million farming households.
These institutions share a common structural problem. They generate intelligence every day. That intelligence does not compound. Staff turn over. Context is lost. The next loan officer starts from zero. The next doctor at the PHC does not know what the last doctor knew.
Figure 5: The institutional memory problem across sectors
COS is designed to solve this problem at institutional scale. From the national headquarters down to the last-mile edge. In 42 languages. Offline-first.
The Last Person First Principle
COS Principle 6 is Last Person First. The framework is not designed for the institution at the top of the hierarchy. It is designed to make the person at the operational edge as capable as the person at the centre.
Figure 6: COS deployment hierarchy
The ASHA worker in rural Odisha should have access to the same quality of clinical decision support as a specialist at AIIMS Delhi. The cooperative bank loan officer in Vidarbha should have access to the same credit risk intelligence as a PSU bank underwriter in Mumbai.
Not the same compute. The same intelligence. Compounded from the same network. Delivered where it is needed, offline if necessary.
Compliance as a Structural Property
COS Principle 3 is Compliance = Structural. Regulatory obligations are not addressed through periodic audits. They are enforced at the operating system layer, before any operation executes.
AdiOS implements this for the Indian regulatory context:
- DPDP Act 2023 data boundary enforcement
- RBI AI guidelines for BFSI deployments
- IRDAI for insurance AI operations
- SEBI algorithmic governance rules
- ABDM FHIR R4 for healthcare data exchange
The distinction matters. Procedural compliance fails at the speed of AI operations. Structural compliance does not. When 10,000 inference calls happen in a day, an audit trail reviewed monthly is not compliance. Enforcement at execution time is.
The Architecture Is Not India-Specific
The COS framework applies to any regulated market where:
- Intelligence must compound inside the institution permanently
- The operational edge must work offline at ultra-low power
- Compliance must be structural, not procedural
- Value must grow as the network of participating institutions grows
Gulf sovereign wealth funds. Southeast Asian central banks. African national health authorities. European enterprises under GDPR. The institutional memory problem is universal. The last-mile problem is universal. The compliance-as-structure requirement is becoming universal.
India is the right starting point because the regulatory urgency here is the most acute, the scale is the most demanding, and the last-mile complexity is the most extreme. Proving COS at this scale is the proof that travels.
The Thesis
AI intelligence is memory, not a service.
When intelligence is a service, the institution rents it. When intelligence is memory, the institution owns it. Every operation, every observation, every validated insight builds the Org Brain. The institution grows smarter with every interaction.
The goal is not to displace the large-scale AI infrastructure that major enterprises already run. The goal is to make circular intelligence accessible to every institution. The PSU bank in Mumbai and the cooperative bank in Manipur. The Apollo hospital in Chennai and the PHC in Bastar. The large agri-cooperative and the farmer producer organisation in Vidarbha.
Intelligence as memory. For every institution. At every scale.
AdiOS Platform Private Limited | Hyderabad, India
adiosplat.io | [email protected]
Originally published on LinkedIn on April 16, 2026.