// DOCUMENTATION

Aegis Refine docs

Aegis is an autonomous dataset refinery. It turns messy, unstructured data into clean ShareGPT / ChatML datasets — governed by the custom Aegis-14B model on NVIDIA DGX Spark, with a human gate on every external spend and a signed audit certificate for every job.

FastAPI · Python PostgreSQL Stripe Aegis-14B · DGX Spark Ed25519 certificates

Quickstart

1
Point Aegis at your data
Give Aegis an https URL to your raw source — Discord logs, PDFs, scraped HTML, or raw JSON — and describe the dataset you want out.
2
Pay the flat fee
$20 via Stripe Checkout starts the job. A verified checkout.session.completed webhook creates the Job record and triggers Aegis-14B triage.
3
Approve any proposed spend
If the agent needs a paid external step (OCR, enrichment), it proposes a SpendTicket. You approve or reject it from the Ops console — nothing outbound runs without that gate.
4
Download dataset + certificate
Get clean JSONL plus a signed audit certificate proving exactly what happened — every action, actor, and authorization.
Start a job ↗

The Aegis loop

Every meaningful action passes through the Aegis-14B governance loop. Aegis-14B runs locally on DGX Spark and scores each job before anything outbound can run — the loop is not infrastructure, it is the product.

triage→ score→ route→ propose→ human_gate→ aar

The human_gate step is the one place a person is required. Everything before it is autonomous scoring; everything after it is the signed record.

Data models

Core MVP entities. These map directly to the PostgreSQL schema and to the JSON the API returns.

▢Jobthe unit of work
iduuidprimary key
user_iduuidowner
statusenumtriage · refining · awaiting_gate · complete · halted
complexity_scorefloat0–1 from the Aegis-14B score node
estimated_costdecimalflat fee + projected spend
input/output_pathstringsource URL and delivered JSONL
▢SpendTicketthe gate
iduuidprimary key
job_iduuidparent job
amountdecimalproposed external spend
descriptionstringreason / rationale
statusenumproposed · approved · rejected · executed
stripe_pi_idstringset once executed (test stub today)
▢AuditLogimmutable
iduuidprimary key
job_iduuidparent job
actionstringpropose_spend · human_approve · execute …
actorstringsystem / agent / human email
detailsjsonbredacted, whitelisted event payload
▢AuditCertificatethe deliverable
iduuidprimary key
job_iduuidparent job
json_pathstringsigned JSON certificate (AAR)
pdf_pathstringrendered PDF
signaturestringEd25519 signature

Spend gate policy

No outbound spend executes without passing the gate. The policy is evaluated on every SpendTicket the agent proposes.

# POLICY-DATA-REFINE-01
if ticket.amount <= job.auto_approve_threshold:
    authorize()              # logged, no human needed
else:
    require_human_approval()  # pauses at gate
always:
    audit_log.append(action, actor, details)
⛉

Honest status: approving a ticket marks the spend authorized (test). Outbound execution is still a test stub — no real funds move yet. Inbound revenue (the $20 Stripe Checkout) runs in Stripe test mode.

Audit certificate

Every completed job produces a signed certificate — the deliverable customers pay for. It is the full AAR record (subject, principal, task, verdict, checks) and verifies independently against the published did:web key. Available as JSON and PDF.

cert_id        string     # cert_7a2b9c4d81
job_id         string     # JOB-2026-00147
issued_at      datetime
ledger         object     # revenue, spend_authorized (test), approvals
decisions[]    array      # action, actor, amount
dataset        object     # rows, quality, pii_masked
signature      string     # ed25519
View a rendered certificate ↗

API

The endpoints below are live today (same-origin; admin routes need an X-Admin-Key). See API_CONTRACT.md for exact request/response shapes.

POST  /jobs/                  # start order → Stripe Checkout
GET   /jobs/:id               # status + gate state + certificate
POST  /jobs/:id/process       # Aegis-14B triage · quality · spend
POST  /admin/gate/:id/approve # human gate (X-Admin-Key)
POST  /jobs/:id/complete      # issue the signed AAR
GET   /jobs/:id/aar           # signed certificate
GET   /.well-known/did.json   # public key for verification
GET   /activity               # redacted live feed

Public REST PREVIEW

A versioned public REST surface for creating jobs, polling status, and approving spend tickets is coming with Team & Volume plans.

POST   /v1/jobs                # create + pay
GET    /v1/jobs/:id            # status + gate state
POST   /v1/spend_tickets/:id/approve
GET    /v1/jobs/:id/certificate