GI
NORTH COUNTY, SAN DIEGO, CA
OPERATIONAL
Garden Insight, Inc.
Sovereign Agentic Intelligence Platform
SOVEREIGN
INTELLIGENCE
INFRASTRUCTURE
Full-stack AI. On-premise. Under control.
Mission
Deliver productive, secure, actionable & governable agent-driven and agent-owned teams, organizations and companies.
LOCATION San Diego
NODES 2 Online · More Soon
VRAM 216 GB Non-Cloud
PROJECTS 19 Core · Rapidly Growing
01 / 10
The Foundation
WHY BUILD
SOVEREIGN AI?
01
Data Sovereignty
Sensitive workloads — client data, proprietary models, civic infrastructure — cannot leave the building. Inference at the frontier is a data liability. You need the model next to the data.
Security First
02
Economics at Scale
Frontier API costs become prohibitive at inference volume. Owning the compute converts per-token rent into amortized capital. The math compounds. At scale, the lab pays back in months, not years.
Cost Control
03
Architectural Ownership
Building entirely on rented intelligence is architectural debt. Sovereign capability means controlling the model, the context, the latency, and the roadmap — not waiting for an API changelog.
Long-term Leverage
02 / 10
Operating Model
THE AGENTIC
ENTERPRISE
PrincipalsHuman · Agentic · Hybrid
Autonomous
Human-Only
Hybrid
◈
Strategists
Vision · Goals · Governance
◉
Domain Experts
Knowledge · Validation · Review
◎
Stakeholders
Approval · Oversight · Direction
INTENT + GOVERNANCE LAYER
AI
Coordination
Orchestration
Task routing · Memory · Context assembly
Human Review Gate
Approve → Execute · Reject → Loop
OPTIONAL
Autonomous Exec
Defined scope · Self-correct · Report
AUTONOMOUS AGENT TIER
Agent
Swarms
Research Swarm
AGENT-R1
AGENT-R2
AGENT-R3
Development Swarm
AGENT-D1
AGENT-D2
AGENT-D3
Operations Swarm
AGENT-O1
AGENT-O2
AGENT-O3
Vision
Horizon
Agentic Sub-Org: R&D
Own goals · Own tools
AUTONOMOUS
Agentic Sub-Org: Ops
Own budget · Own roadmap
AUTONOMOUS
Agentic Sub-Org: Civic
Public benefit mission
AUTONOMOUS
03 / 10
Architecture
CLICK A LAYER
TO EXPLORE
Applications Layer
Human Intent
→
Application Harness
→
⊕ Human Review Gate
✓ Approved
Human validated
→
Output
⚡ Autonomous Exec
Scoped permission
→
Output + Report
Human-in-the-loop gates decide when agents act independently vs. when approval is required. Every output traces back to human intent.
Intelligence Layer
Request
→
Intent Router
→
Response Assembly
→
Output
80–90%
Local Inference
Qwen3 · Cogito
On-premise · Low latency
Full data sovereignty
10%
Frontier Cloud
Claude · GPT · Gemini
Capability ceiling tasks
Reserved for novel work
A deliberate architectural boundary. Local handles volume; frontier handles ceiling tasks.
Memory & Data Layer
01
Classify & Route Request
Context Assembler
02
Semantic Vector Search
pgvector
03
Graph Relationship Query
Apache AGE
04
Cache & Session State
Redis
05
Assemble & Rank Context
Assembler Pipeline
→
Inject into Agent Context Window
Output
Platform Layer
Proxmox VE 9.1
Hypervisor
→
k3s
Container Platform
→
Tekton + ArgoCD
CI/CD · GitOps
FreeIPA
Central Auth
OpenBao
Secret Mgmt
Tailscale
Zero-Trust
Identity-first platform design — no one-off users, no secrets in runbooks.
Conceptual Compute Roles
⬡
Inference Engine
Primary GPU compute. Large model serving. Full VRAM for context window size and throughput.
⬢
Swarm Executor
Parallelism tier. Multiple agents concurrently, isolated workloads, GPU partitioning via SR-IOV.
◈
Control Plane
Orchestration, memory graph, CI/CD pipelines, identity services, system state.
◉
Edge Coordinator
Lightweight inference, API routing, cloud burst gateway, UI serving, task dispatch.
Any node can serve any role. Composable by design — ADR-014.
04 / 10
Development Lifecycle
AGENTIC
WORKFLOW
⊕ Context Context
WHO
Entity · Role · Org
WHY
Business intent · OKR link
WHAT
Artifact type · Gov class
HOW
Quality gates · Criteria
WHERE
Scope · Permission boundary
Backlog
◉ Human
Business intent defined
Scope & priority set
OKR linkage captured
AUTO
CC ASSEMBLE
Context Ready
Human + Agent
CC assembled & scored
ADRs resolved
Artifacts defined
Gov class set
AUTO
IF SCORE OK
In Progress
⬡ Agent
Autonomous execution
Swarm coordination
Self-correction loops
Own pace — no standups
AGENT
SELF-VERIFY
Tests Passing
⬡ Agent
Automated validation
Artifact verification
Traceability check
CONFIG
ORG-DEFINED
Human Review
◉ Human
Governance gate
NL-configured policy
Bypassable by tier
GATE
APPROVE
Done
✓ Complete
Artifact published
Versioned in git
Discoverable
Audit trail sealed
Agents Work UnblockedBetween gates, agents run at machine pace — no standups, no sprint ceremonies.
NL-Configured GatesPlain-language policies become enforced gates automatically.
Flexible by TierCompany, org, and team each define their own gate thresholds.
Every Artifact is GovernedAuto-classified, versionable, auditable, discoverable from intent to deploy.
05 / 10
Governance & Traceability
GOVERNANCE-FORWARD
ENTERPRISE WORK
Tier 01 · Company
Enterprise Governance
GI Classification HeadersADR ComplianceSecurity PostureAudit Retention
Tier 02 · Organization
Org-Level Policy
Custom Gate RulesApproval ChainsRole-Based FlowsDomain Boundaries
Tier 03 · Team
Team Autonomy Config
Agent Autonomy LevelReview FrequencyBypass ThresholdsSwarm Scope
Natural Language Policy Interface
"All changes to GI-RESTRICTED artifacts in the security domain require human review and dual approval before merging."
Parsed → gate rule created → version-pinned → enforced at CI gate → included in audit log
"Allow the ops swarm to auto-deploy infrastructure changes under 50 lines if all tests pass."
Parsed → autonomy policy set → scoped to ops swarm → logged as team policy v1.2
End-to-End Traceability
Business Outcome
Biz Layer
↓
OKR / Strategic Goal
Planning
↓
Epic → Story → Task
Work Mgmt
↓
Agent Action Log
Execution
↓
Artifact (code · doc · config)
Output
↓
Deployment · Integration Test
Operations
↓
Metric · Business Outcome ↩
Feedback
Versionable
Git-backed, diff-able, semantic versioning on all work products including policies and agent configs
Discoverable
Semantic search via pgvector. Every artifact tagged by domain, team, gov class, and lifecycle state
Auditable
Full provenance from intent to deployment. Human and agent actions both logged with immutable chain
06 / 10
Technology Radar
TECHNOLOGY
LANDSCAPE
EMERGING INVEST. STANDARD INTELLIGENCE PLATFORM INFRA DATA Proxmox vLLM PostgreSQL Ubuntu k3s FreeIPA Tailscale Qwen3 OpenWebUI pgvector Redis Tekton ArgoCD OpenBao SR-IOV Cogito 70B Apache AGE RKE2
Compute
Proxmox VECUDA 12.9VFIO/IOMMUk3sSR-IOV
Inference
vLLMQwen3 FamilyOpen WebUICogito 70BBenchmark Harness
Data & Memory
PostgreSQL 16pgvectorRedisApache AGEGraph Memory
Identity
TailscaleBIND9FreeIPAOpenBao
CI/CD & Orch
TektonArgoCDRKE2Karpenter
Frontier AI
Claude APIOpenAI APIGemini APICoreWeave
Knowledge
NotionPrompt Librarypgvector+AGE KG
Standard
Emerging
Investigating
07 / 10
Research & Development
R&D
PRIORITIES
01
Agentic Memory Architecture
Design and deploy a hybrid semantic + graph memory system that gives agents persistent, queryable context across sessions. Enables true multi-session reasoning, relationship tracking, and institutional memory — without starting from scratch every call.
pgvectorApache AGEContext AssemblyGraph Reasoning
02
Multi-Model Inference Pipeline
Build a rigorous benchmark harness and model leaderboard for evaluating local models across coding, reasoning, and instruction-following tasks. Enables confident model selection, continuous evaluation, and regression detection as models evolve.
Benchmark HarnessModel LeaderboardGPU UtilizationInference Perf
03
Autonomous Agent Swarms
Research and implement agent swarm coordination protocols: parallel task execution, inter-agent communication, self-correction loops, and human escalation gates. Builds toward fully autonomous sub-organizations with defined goals, budgets, and accountability.
Swarm ProtocolsSelf-CorrectionGovernance GatesAgent Autonomy
04
Civic AI & Public Benefit
Applying sovereign AI infrastructure to public-good use cases is a mission priority. Enterprise deployments are the proving ground for the stack — but the same knowledge and capabilities belong in service to social enterprise and nonprofit organizations. We endeavour to give back with the tools and knowledge that we are building.
Social EnterpriseNonprofitPublic BenefitOpen Data
08 / 10
Continuous Autonomous Research
AGENTIC
RESEARCH FLEET
4
Active
4
Ongoing
4
Queued
6
Domains
Research agents run continuously and autonomously — scanning literature, synthesizing findings, surfacing insights to the knowledge graph. Human review triggered when confidence thresholds are exceeded or novel findings require validation.
Active
Agentic Context Architecture
Hybrid pgvector + graph memory systems, 6-step context assembly pipelines, cross-session agent continuity and institutional memory accumulation patterns.
Memory · Architecture
Active
Multi-Model Eval Framework
Automated benchmarking across coding, reasoning, and instruction domains. Capability-cost mapping, model selection heuristics, regression detection at model update cadence.
Intelligence · Eval
Active
Swarm Coordination Protocols
Inter-agent messaging standards, parallel task decomposition strategies, consensus and conflict resolution mechanisms for heterogeneous agent populations.
Agents · Orchestration
Active
NL Policy Translation Engine
Parsing natural language governance rules into machine-enforceable gate policies. Version control, rollback, and conflict detection for overlapping policy definitions.
Governance · Platform
Ongoing
SR-IOV GPU Partitioning
GPU virtualization patterns for isolated per-agent VRAM allocation. Workload scheduling, memory bandwidth contention modeling, QoS guarantees across agent swarms.
Infrastructure · Compute
Ongoing
Hybrid Inference Routing
Dynamic local/frontier routing based on real-time cost, capability, latency, and classification constraints. Auto-learning routing policies from outcome signals.
Intelligence · Cost
Ongoing
Civic AI · Public Sector
Government meeting transcription, entity recognition, public record analysis, and structured output generation. Accuracy benchmarking against human transcriptionists.
Applications · Civic
Ongoing
Knowledge Graph Evolution
Tracking entity relationships and knowledge state over time. Detecting semantic drift, pruning stale context, and maintaining coherence across long-running agent sessions.
Memory · Data
Queued
Autonomous Org Bootstrap
Protocols for spinning up agent-owned sub-organizations with defined goals, scoped budgets, accountability structures, and human escalation pathways.
Agents · Autonomy
Queued
Multi-Modal Context Integration
Incorporating audio, visual, and structured data streams into the context assembly pipeline. Unified context scoring across modalities for heterogeneous agent inputs.
Intelligence · Multimodal
Queued
Cross-Org Knowledge Federation
Sharing institutional memory across organizational boundaries with appropriate access controls, classification enforcement, and provenance tracking.
Platform · Enterprise
Queued
Agent Self-Improvement Loops
Structured experimentation frameworks enabling agents to improve performance on defined metrics through controlled self-modification within governance boundaries.
Research · Meta-AI
Research pipeline is continuously updated by the autonomous research swarm. Findings are published to the internal knowledge graph and surfaced for human review on a rolling basis.
09 / 10
The Road Ahead
THE
HORIZON
The goal is not to use AI.
The goal is to own the capability — at every layer of the stack.
A sovereign AI lab is not an IT project. It is an institutional capability — the ability to reason, build, and operate at machine scale, under your own authority, with your own data, on your own terms.

Garden Insight is building that capability from the ground up, one composable layer at a time.
Complete · Jan–Mar 2026
Compute + DNS Foundation
Primary inference node online · 96 GB VRAM live · Internal DNS namespace established
Active · Apr–May 2026
Identity, Auth & Storage
FreeIPA central directory · OpenBao secrets · DHCP · Atlas-Compute fully integrated
May–Jun 2026
Platform & CI/CD
k3s cluster · Tekton pipelines · ArgoCD GitOps · Agentic dev workflow tooling live
Jun 2026
Agentic Memory + Swarm Platform
pgvector + AGE graph memory · Agent swarm tier live · Full agentic pipeline operational
Late Jun 2026
The Agentic Enterprise
Autonomous sub-organizations · 216+ GB non-cloud VRAM · Sovereign memory · Governance-forward workflows at scale
Jul 2026 →
Classification as Architecture · Precision Context Assembly · Agent Clean Room
Data-flow governance, not just action governance. An Agent Clean Room governs what leaves. Auditable evidence at every decision point. The architecture forces classification and privacy compliance — policy documents alone cannot.
10 / 10