// HOW IT WORKS

How Aegis Refine runs an agent-operated data business.

Customer data goes in, Stripe Checkout earns the capped quote, Hermes Agent operates the job through NemoClaw/OpenShell when enabled, NVIDIA models make the route and safety decisions, Stripe Connect handles approved spend, and the customer receives a signed dataset certificate.

// FULL SYSTEM MAP

One workflow, four proof layers.

The app is not just a form around a model. It is a bounded business loop: earn through Stripe Checkout, operate through Hermes Agent, spend through a capped Stripe Connect Transfer, and prove the result with an AAR certificate.

flowchart LR
  Customer["Customer"] --> Order["New dataset job"]
  Order --> Quote["Aegis quote
flat capped price"] Quote --> Checkout["Stripe Checkout
earn rail"] Checkout --> Job["Paid Job"] Job --> Sandbox["NemoClaw/OpenShell
aegis-hermes sandbox"] Sandbox --> Hermes["Hermes Agent
aegis-refine skill"] Hermes --> Ultra["Nemotron 3 Ultra
operations brain"] Hermes --> Aegis["Aegis-14B
data governance"] Hermes --> Safety["Nemotron 3.5
content safety gate"] Aegis --> Work["Refine or synthesize
on DGX Spark"] Safety --> Work Ultra --> Gate{"Need paid tool
within cap?"} Gate -->|No| Work Gate -->|Yes| Spend["SpendTicket
human approval"] Spend --> Transfer["Stripe Connect Transfer
AINode compute"] Transfer --> Verify["Backend verifies
tr id, cap, destination"] Verify --> Work Work --> Cert["Signed AAR certificate
dataset + receipts"] Cert --> Customer
// NEMOCLAW / OPENSHELL

Hermes can run inside a live NemoClaw/OpenShell sandbox.

The Dell R750 hosts a sandbox named aegis-hermes. The private operator bridge sends the same bounded job payload into that sandbox with nemohermes exec, where Hermes loads the aegis-refine skill. Receipts record the runtime mode, sandbox name, and runtime-configured NVIDIA inference model.

CODE PATH
hermes/operator_bridge.py switches between local, NemoClaw, and OpenShell modes with HERMES_OPERATOR_RUNTIME.
SKILL
hermes/aegis-refine/SKILL.md defines the operator task schema and fail-closed receipt contract.
LIVE SANDBOX
aegis-hermes runs on the Dell R750 through NemoClaw/OpenShell for the demo operator path.
GPU NOTE
The host detects the NVIDIA A40; sandbox GPU passthrough is documented as follow-up pending NVIDIA Container Toolkit/CDI setup.
flowchart TB
  Backend["Aegis backend
paid job payload"] --> Bridge["Hermes operator bridge
auth + bounded prompt"] Bridge --> Nemo["NemoClaw/OpenShell
aegis-hermes sandbox"] Nemo --> Hermes["Hermes Agent
--skills aegis-refine"] Hermes --> Broker["inference.local
host-managed broker"] Broker --> Nvidia["NVIDIA Endpoints
Nemotron route"] Hermes --> Receipt["operator receipt
runtime + model + route"] Receipt --> Backend
// HERMES + NVIDIA ROLES

Each model has one job.

The receipt names the actual model used. Ultra is the primary operations brain; Nano is only a latency fallback when recorded as such. If the content-safety model did not inspect raw evidence, the receipt says metadata-only or pending.

OPERATOR
Hermes Agent loads the Aegis Refine skill and executes the business process.
SANDBOX RUNTIME
NemoClaw/OpenShell runs the Hermes operator path when HERMES_OPERATOR_RUNTIME=nemoclaw.
DATA GOVERNANCE
Aegis-14B, trained from Hermes 14B, judges dataset quality, risk, and refinement route.
OPERATIONS BRAIN
nvidia/nemotron-3-ultra-550b-a55b makes routing, cap, and spend decisions.
SAFETY GATE
nvidia/nemotron-3.5-content-safety is the PII and unsafe-content review layer when evidence is available.
flowchart TB
  Runtime["NemoClaw/OpenShell
sandbox runtime"] --> Hermes["Hermes Agent
operator runtime"] Hermes --> Skill["aegis-refine skill"] Skill --> Ultra["Nemotron 3 Ultra
routing + cap decisions"] Skill --> Aegis["Aegis-14B
dataset governance"] Skill --> Safety["Nemotron 3.5 Content Safety
PII and unsafe-content gate"] Skill --> Stripe["Stripe skill / transfer helper
approved spend only"] Ultra --> Receipt["operator receipt"] Aegis --> Receipt Safety --> Receipt Stripe --> Receipt
// EARN + SPEND RAILS

Stripe is used twice, and the two flows are deliberately separate.

Checkout is the earn rail: the customer pays the capped quote. Connect Transfer is the spend rail: Hermes Agent spends only after the cap gate approves it, and the backend independently verifies the Stripe transfer before recording the ticket as executed.

sequenceDiagram
  participant Buyer
  participant App as Aegis Backend
  participant Stripe as Stripe
  participant Hermes as Hermes Agent
  participant Vendor as AINode Vendor
  Buyer->>App: Request quote
  App-->>Buyer: Flat capped price
  Buyer->>Stripe: Pay Checkout
  Stripe-->>App: checkout.session.completed
  App->>Hermes: Dispatch paid job
  Hermes->>App: SpendTicket proposed if needed
  Buyer->>App: Approve spend
  App->>Hermes: spend_approved with cap
  Hermes->>Stripe: Create Connect Transfer
  Stripe-->>Hermes: tr_ transfer id
  Hermes-->>App: Return tr_ id
  App->>Stripe: Retrieve transfer
  App->>App: Check cap and destination
  App-->>Buyer: Verified spend receipt
  Stripe-->>Vendor: Transfer to AINode compute
      
flowchart TB
  Start["Agent claims spend"] --> HasId{"Returned tr_ id?"}
  HasId -->|No| Queue["temporarily queued"]
  HasId -->|Yes| Retrieve{"Stripe retrieve succeeds?"}
  Retrieve -->|No| Queue
  Retrieve -->|Yes| Cap{"Amount <= approved cap?"}
  Cap -->|No| Queue
  Cap -->|Yes| Dest{"Destination is AINode vendor?"}
  Dest -->|No| Queue
  Dest -->|Yes| Execute["Mark SpendTicket executed
record livemode + transfer id"]
// FAIL-CLOSED RULES

No fake fallback, no synthetic payment IDs.

If any proof step fails, the job does not pretend. It moves to temporarily queued or asks an operator. That is the point of Aegis: an agent can act, but every economic claim has to verify.

VERIFY AGAINST STRIPE
The backend retrieves the transfer object from Stripe before recording spend.
CAP RESPECTED
The verified transfer amount must be less than or equal to the approved cap.
DESTINATION LOCKED
The destination must match the configured AINode compute vendor account.
HONEST MODE
Receipts store Stripe livemode from the Stripe object, so sandbox demos are labeled honestly.
// CUSTOMER LOOP

What the buyer experiences.

The public workflow stays simple: submit data, accept the capped quote, let the agent operate, and download the refined dataset plus proof.

// THE GUARANTEE

A cap you approve. Then it runs on its own.

The quote is flat and capped before any work starts. Accepting it pre-authorizes the agent to run autonomously inside that budget — it can’t bill you past the cap. The only time it stops to ask is if finishing a job would exceed the cap you approved; then you approve a top-up or keep your cap. Either way, no surprise bill.

See a signed certificate →
FLAT CAPPED QUOTE
YOUR CAP
$20.00
RUNS WITHIN CAP
SPENT $4
PINGED ONLY ON A CAP OVERRUN

Accept a cap, then it runs on its own

$
You accept the quote

Pay the flat, capped price via Stripe Checkout. The cap becomes the agent’s budget.

Agent runs within budget

Refines or synthesizes autonomously on DGX Spark — and meters its own real spend against the cap.

Pinged only on overrun

If finishing would exceed your cap, it asks — approve a top-up or keep your cap. Otherwise, no interruptions.

Bulk work runs on NVIDIA DGX Spark and Hermes can run through NemoClaw/OpenShell

The custom Aegis-14B model governs every job from local inference. The Hermes operator path can be sandboxed through NemoClaw/OpenShell, while paid external APIs are reserved for cases that justify them and only within the budget you approved.

LOCAL-FIRST · SPARK 1