Platform Gateway
- identity · tenant
- request context
- authn / authz context
AI platform architecture / current boundary
Building toward a Governed Enterprise AI Platform
I built these systems independently and tested each one against its own evidence. This page shows the larger architecture I’m evolving them toward.
From trusted organizational knowledge, to structured-data reasoning, to controlled business action.
01Evidenceimplemented systems first
02Architecturerelationships earned by evidence
03Integrationfuture work stays labeled
WHERE THE SYSTEMS ARE GOING
How the systems are evolving together.
The proven nodes are existing, independently tested systems. Building and evolving nodes show active work or the direction of integration.
KNOW
What does the organization know?
UNDERSTAND
What does the structured data tell us?
ACT
What should the system do?
SHARED MODEL RUNTIME
MODEL LIFECYCLEpromotion / rollback
existing relationship / implementation target integration
Expand full architecturethe target system, with what is built marked
The graph above shows what exists and how it connects today. This is the architecture those systems are being built toward. Every box carries its own status, because without that the target reads as though it were already finished.
Request entry
Capability Router
NEXT- which capability handles this request
Capability planes
Each plane proposes; the deterministic controls inside it decide.
Knowledge plane
PROVEN- retrieval + reranking
- evidence construction
- citation validation
- fail-closed abstention
Data plane
BUILDING- schema retrieval
- SQL generation
- AST validation
- table / column ACL
- read-only execution
Action plane
PROVEN- target resolution
- policy
- confirmation
- revalidation
- idempotent execution
Shared model runtime
Model choice never changes permissions.
Adaptive Model Router
EVOLVING- task · capability · complexity aware
- quality · latency · cost aware
- privacy and local-only aware
- health · fallback · escalation
Model tiers
NEXT- specialist SLM
- general reasoning LLM
- vision and multimodal
Inference providers
- local open models
- OpenAI-compatible APIs
Model lifecycle
A candidate reaches the router only through an evaluation it passed.
FineForge
NEXT- QLoRA / PEFT
- dataset preparation
- resource profiling
Offline evaluation
NEXT- base vs fine-tuned
- quality · latency · VRAM
- generalization
ModelOps Control Plane
PROVEN- canary 10 → 25 → 50 → 100
- quality and reliability gates
- delayed ground truth
- promotion · rollback · reconciliation
Model registry
NEXT- versions and lineage
- feeds the router
Cross-cutting governance
Shared contracts and central governance over independent services, not one monolith.
Enterprise Context
NEXT- one entity across planes
- data · knowledge · metrics · actions
- role / tenant / scope
Capability Gateway
NEXT- typed capabilities
- REST / OpenAPI · MCP
- databases · knowledge · business systems
Unified Registry
NEXT- models · agents · knowledge bases
- entities · tools · policies
- eval suites · deployments
Control Plane
NEXT- identity · policy · lifecycle
- evaluation · deployment state
- observability · audit · cost
One trace across every plane
knowledge.rag
- retrieve
- rerank
- evidence.build
- model.route
- validate
decision.sql
- schema.retrieve
- sql.generate
- sql.parse
- policy.evaluate
- query.execute
agent.runtime
- proposal
- target.resolve
- policy.evaluate
- confirmation
- tool.execute
Proof before vision
Implemented systems.
Three independent subsystems, each with a different failure surface and an explicit evidence trail.
Status is a claim boundary, not a project-management label.
PROVENImplemented independently and backed by explicit evidence.
BUILDINGActive implementation.
EVOLVINGA working subsystem exists; its role or abstraction is being generalized.
NEXTNot implemented yet.
KNOW · Trusted organizational knowledge
Knowledge Base RAG
Evidence-backed access to unstructured organizational knowledge.
Engineering decision
Tenant ACL and evidence validation stay server-owned; unsupported answers fail closed.
Current evidence
- 95.9% candidate evidence recall
- 220-query frozen reranker benchmark
- 0 citation contract failures
Stack / boundaries
- Qdrant
- Qwen3 embeddings
- BM25
- RRF
- BGE reranking
- tenant ACL
- citation validation
- fail-closed abstention
- OpenTelemetry
ACT · Agent Runtime foundation
Agent Runtime
Controlled business actions where the model proposes intent but deterministic software owns execution authority.
Engineering decision
The LLM never receives mutation authority.
Current evidence
- 540/540 measured deterministic safety attempts
- 0 unsafe executable survivors
- 0 unsafe executions
Stack / boundaries
- LangGraph
- FastAPI
- PostgreSQL
- SQLAlchemy
- typed tools
- deterministic policy
- confirmation
- revalidation
- idempotency
- OpenTelemetry
MODEL LIFECYCLE · Evidence-based model lifecycle management
ModelOps Control Plane
Canary delivery and promotion control where candidates earn traffic through explicit evidence gates.
Engineering decision
A candidate earns production traffic by satisfying predefined evidence gates, not by existing.
Current evidence
- 10% → 25% → 50% → 100% automated canary progression
- PASS / FAIL / INCONCLUSIVE verdicts with automatic rollback paths
- 279 backend tests · ~91% statement coverage
Stack / boundaries
- FastAPI
- SQLAlchemy
- canary rollout
- quality + reliability gates
- delayed ground truth
- automatic rollback
- reconciliation
- Docker Compose
Model lifecycle / model optimization
Specialization has to earn its place.
FineForge is intentionally shown as NEXT here. The question is not whether a model can be fine-tuned; it is whether specialization actually beats the base model.
01Base Model
02FineForge / QLoRA
03Offline Evaluation
04Candidate
05ModelOps
06Canary
07Registry
08Adaptive Model Router
NEXT
FineForge / QLoRA
Potential evaluation: task quality, SQL execution accuracy where relevant, retrieval/generation quality, latency, throughput, VRAM, cost, and generalization.
Shared engineering principles
Engineering principles.
The platform direction is coherent because the boundaries repeat, even when the subsystems do different work.
Models provide intelligence. Deterministic systems provide authority.
Evidence decides what gets promoted.
Evaluation gates are defined before looking at the result where applicable. A candidate can improve some metrics and still be rejected. Rejected experiments remain evidence instead of being rewritten as success.
Read the frozen reranker decision (opens in a new tab)Failures stay attributable.
retrieval miss ≠ reranker loss ≠ generation error ≠ validator error ≠ policy failure ≠ execution failure
Aggregate “AI accuracy” is not enough to decide which subsystem needs to change.
Integration roadmap
Integration second. Platform last.
These are target integrations, not current deployment claims. The sequence starts with a second real data consumer and grows the contracts only when the next boundary is earned.
- BUILDING
DecisionSQL
Structured data plane
- NEXT
Enterprise Context V0
One entity consumed across planes
- NEXT
Cross-plane integration
Knowledge, data, and action contracts
- NEXT
Shared capability contracts
Stable boundaries between bounded services
- NEXT
Unified Registry
Models, agents, knowledge, data, tools, policies, evaluations, deployments
- NEXT
Unified Control Plane
Common identity, configuration, policy, and lifecycle
NEXT
Enterprise Context V0 starts with Order.
Enterprise Context starts only after DecisionSQL creates a second real data consumer, avoiding a speculative ontology layer.
The first success criterion is simple: the same Order entity is actually consumed by RAG, SQL, and Agent subsystems.
Order
data: commerce.ordersknowledge: order-policy · refund-policymetrics: order_value · refund_rateactions: cancel_order · refund_orderKnowledge Base RAGDecisionSQLAgent RuntimeOrder
NEXT
Unified Registry
Models · Agents · Knowledge Bases · Data Sources · Tools · Connectors · Policies · Evaluation Suites · Deployments
NEXT
Unified Control Plane
Common identity, configuration, policy, and lifecycle across independent bounded services.
Independent bounded services+shared contracts+common control plane
Not one giant repo containing every project.
Built independently. Measured explicitly. Now being connected deliberately.