mindpool · Enterprise-Owned AI Infrastructure
$ cat ./enterprise

Own the intelligence your enterprise runs on.

Watch · The Intelligence Layer · 73s

Mindpool builds infrastructure for enterprises to own and compound the intelligence their work creates. Open models, open-source components, and systems running in your environment keep that learning under your control. Start with Bowline — the open-source intelligence layer running in your environment today.

the outcome

Ownership is the foundation. Compounding is the return.

Every evaluation, correction, decision, and workflow creates knowledge about how your organization operates. That intelligence should accumulate inside a boundary you control — even when some work runs on rented models.

$ cat ./three-tenets

Cost. Control. Data sovereignty.

AI becomes yours when you control the economics, the decisions, and the data. These are the three conditions of ownership.

Cost
Economics you can act on.

AI spend shouldn't be a bill you read after the fact. See the economics first — measured on your own traffic, against supply you own or rent, open weights in your VPC included — and decide with evidence.

Control
Authority you grant — never authority it takes.

No black-box routing, no vendor console deciding for you. Decisions are recorded, evidence is verifiable, and anything that touches live traffic is an explicit step an operator takes — with a way back.

Data sovereignty
Your traffic, your evidence, your security domain.

Models, weights, and data stay in your environment. What leaves crosses a gate you govern — and the evidence of what happened stays on infrastructure you hold.

$ read ./essays/the-intelligence-layer
category thesis · essay 01

Every company will own more of its AI. Who decides where the work belongs?

The enterprise AI portfolio now spans owned hardware, open weights in a VPC, managed endpoints, and public APIs. The missing layer is judgment across them — with evidence the enterprise keeps.

› read The Intelligence Layer
$ ls ./products

Two products. Two layers. One architecture you own.

intelligence layer · open source · Apache-2.0
Bowline

Bowline decides where each class of AI work belongs and puts every decision on the record. It runs in your environment, watches in shadow mode first, and takes authority only when you grant it. Routers move requests; Bowline decides where work belongs.

execution boundary · open source · Apache-2.0 · now part of Mindpool
NeuronEdge Enclave

A governed sandbox for every AI agent. Each workspace runs in its own microVM — separate kernel, deny-by-default egress, signed audit events — so agent code executes inside a boundary you set, in your environment. A hardware-attested confidential profile is in preview.

$ cat ./architecture

A reference architecture for AI your organization owns

At the top, the value layer — the agents and workflows your organization owns. NeuronEdge Enclave's execution boundary — now part of Mindpool — gives every agent workspace its own isolated kernel and governed egress. Bowline's intelligence layer decides where each class of work belongs — owned supply first, rented only by policy, every decision on the record. Beneath it, the foundation: five fixed infrastructure contracts replace the rented, API-metered stack with sovereign components in your environment. The architecture stays fixed; implementations remain swappable — v1 shipped today, v2 on the roadmap.

The value layer · agent harness

Where the work happens — and where your advantage compounds. Agents and enterprise workflows: engineered by your team or delivered alongside you. This layer is yours to own.

Orchestration
the agent loop — single & multi-agent
Tools & Effectors
MCP tools, invoked under policy
Memory & Context
RAG + working and durable memory
Permissions
side-effect classes + approval ladder
Trajectories
every run replayable and inspectable
Evaluation
measure how an agent fails before it ships
NeuronEdge Enclave · the execution boundaryopen source · Apache-2.0 · now part of Mindpool
v0.1 · standard profile available today

Where agent code actually runs. Every workspace gets its own kernel inside a microVM — an isolation and egress boundary the agents above cannot cross, with audit events an operator can verify.

agent execution — isolate · govern egress · account
replaces shared-kernel containers · uncontained agents
one microVM kernel per workspace · deny-by-default egress · signed audit events
confidential profile — hardware-attested execution (AMD SEV-SNP) · in preview
Bowline · the intelligence layeropen source · Apache-2.0 · runs in your environment
v0.1 · available today

Which intelligence does each task deserve? Bowline decides — owned supply first, rented only when policy allows and the value is real — and puts every decision on the record.

task distribution — classify · gate · place · account
replaces hard-coded model picks · ungoverned AI spend
shadow mode · customer-quality canaries · modeled economics · explicit per-workload enforcement
task decomposition · expanded supply profiles · capability registry feed
rented intelligence
frontier APIs · VPC-hosted models
egress only through Bowline — by policy, accounted
The foundationSovereign, on your hardware, and built to run itself — so your team works in the value layer, not the infrastructure beneath it.
Operated by Mindpool· trust economy — integrated, attested
Economic plane
meter · attest · settle
replaces per-token metered API billing
v1 · shipped
usage ledger + attested benchmarks · per-tenant, peer-tagged usage · billed to your org as a service
v2 · roadmap
metered billing · settled to providers · decentralized compute market (preserved option, not roadmap)
Control plane
reconcile desired-state · schedule · route
replaces the vendor SaaS console
v1 · shipped
operator console — live telemetry · desired-state, Mindpool-run today
v2 · roadmap
hosted control plane · Kubernetes operator · self-serve · SLAs
sovereignty boundary · the stack stays in your environment · metered in place, attested before it crosses
On your hardware · your premises· composable — swap any backend or silicon
Node plane
agent API — telemetry ↑ / desired-state ↓
replaces the vendor SDK / serverless function
v1 · shipped
node agent — streaming API over HTTP + SSE · the same agent on a rig you own or a cloud box launched into the fleet (Azure-validated, torn down clean)
v2 · roadmap
same API (forward-compatible) + autoscale signals
Runtime plane
backend interface — Run / Serve / Finetune / Eval / Quantize / Burst
replaces the hosted inference API
v1 · shipped
llama.cpp-first · MLX · vLLM — pinned, verified, auto-tuned per platform · LoRA finetune (MLX + CUDA live-verified · ROCm hermetic-only) + on-box eval · tenant host — multi-tenant GPU serving: exclusive scheduling, per-tenant peer-tagged usage (validated on Azure T4) · burst jobs run on hardware you own (SkyPilot) or rented managed cloud (HF Jobs)
v2 · roadmap
TensorRT-LLM · NIM · SGLang (adapters built — certification pending) · routing · autoscale
Hardware substrate
capability profile — fit · tok/s
replaces rented GPU / cloud instances
v1 · shipped
AMD (first) + Apple Silicon certified · profiled & benchmarked on the real box
v2 · roadmap
NVIDIA · OEM — matrix expands
Cross-cutting spine· threads every plane, both sides of the boundary
Identity & Trust
signed node identity · entitlement
replaces vendor API keys + IAM
v1 · shipped
SIWE/EIP-191 node identity · on-chain entitlement · bearer-token tenancy
v2 · roadmap
mTLS · signed desired-state · counter-signed attestation (third-party verifiable)
Governance & Supply-chain
provenance · signed artifacts · policy
replaces unsigned downloads · mystery deps
v1 · shipped
signed artifact registry + operator trust store · install-provenance ledger (the SBOM seed) · sha256-pinned models
v2 · roadmap
cosign/OCI + transparency log · SBOM export · org policy engine · SSO/SAML → RBAC

Implementations are dated, evidence-backed picks — swappable behind a fixed contract, re-evaluated, not re-architected. That is what makes no lock-in structural.

The spine rows are cross-cutting, not a sixth plane: signed identity and supply-chain governance run through every plane, on both sides of the boundary.

Roadmap items are directional, not commitments, and subject to change. Today's offering is services-led — Mindpool operates the stack on your hardware.

Put enterprise AI under your control.

Start with Bowline's evidence on your own traffic, or start the conversation about the stack your organization needs to own.