The Sovereign AI Stack Layer by Layer
Most AI platforms ask you to send your data to their cloud. Ooumph sends the model the other way. Your records stay inside your walls, under your control, governed by the policies you already operate to.
Three layers do the work. This page walks through each one: what it is responsible for, and what it guarantees. For specifications, integration APIs, and deployment matrices, see the architecture reference.
Federated Learning Model
The infrastructure layer. The model travels to your secure environment, and training, inference, and learning all happen inside your perimeter. Raw records stay at their origin.
Federated Learning Model (FLM)
InfrastructureThe foundation layer. Training and inference both run inside your institutional perimeter, and the only thing exchanged across the boundary is a federated gradient update. That is a physical property of how the system is wired, so there is no setting that can override it.
- The model is deployed into your environment. You control the weights and the update schedule
- Runs air-gapped where the environment demands it
- During training, only gradient updates cross the boundary
- Because the data never moves, DPDP and GDPR obligations are satisfied by the deployment itself
- Supports thousands of concurrent users across departments
- Sovereignty applies to every query and every training cycle
Federated Thinking Model
Several specialised models work the same institutional problem at once. Where their outputs conflict, the applicable statute decides. No centralised data pooling, and no single point of failure for reasoning.
Federated Thinking Model (FTM)
IntelligenceWhere FLM protects data, FTM produces insight. Several small models, each tuned to a specific institutional function, work the same problem in parallel. Where their outputs conflict, the applicable statute decides the answer rather than a majority vote, so the same inputs reconcile the same way every time.
- Small Language Models, each fine-tuned to a departmental function and its vocabulary
- Reasoning that respects standing orders, procedural hierarchy, and the exceptions
- Conflicting model outputs resolved by statutory precedence rather than by majority vote
- Edge inference fast enough for field and counter use
- Every step of a multi-step task stays readable afterwards
- Departments share intelligence without sharing records
Blockchain Audit Layer
The trust layer. Every interaction writes an immutable audit entry, which is what makes an AI-assisted determination defensible long after the officer who made it has moved on.
Blockchain Audit Layer
TrustAccountability is not optional for public institutions. The Blockchain Audit Layer records every inference, every operator action, and every model version in an append-only ledger. Entries are cryptographically chained: any modification is computationally detectable.
- An append-only record of every query, output, and operator action
- Each entry is cryptographically chained to the one before it via SHA-256
- Entries structured for RTI Section 4 disclosure
- The complete decision pathway can be reconstructed on demand
- Determinations hold up in judicial and administrative proceedings
- Hash-chain validation and anomaly alerts surface tampering
The Engine Underneath All Three Layers
Each layer above describes a responsibility. The OSIE™ inference engine is what actually executes them.
When a query arrives, OSIE picks the specialised model that should handle it, checks the request against the policies in force for that department, runs the inference, and writes the audit entry before the answer is returned. All three layers depend on it, which is why technical evaluators tend to start here.
Latency, throughput, scaling behaviour, encryption standards, key management, and access control are documented in full in the architecture reference.
Read the OSIE SpecificationsWhere the Stack Can Run
The three layers stay the same in all four topologies. What changes is how much of the infrastructure you operate yourself.
On-Premise
All three layers run in your data centre. Nothing about the stack changes; you simply own every component of it.
Air-Gapped
The same three layers, with no external connectivity. Model updates arrive on authenticated offline media.
Hybrid
FLM stays on-premise with the sensitive workloads. Less sensitive inference can route to your approved cloud, with policy deciding placement.
Managed
Ooumph operates the stack under a strict SLA, for departments without a dedicated infrastructure team.
See the Stack in Action
An hour with our engineering team, working through the layers against your own deployment constraints.