AUREX/OX AI intelligence, evidence and human control
Owner Studio / audit workbenchAUREX is the developmental AI intelligence system for multi-model analysis, evidence checking and governed decisions. OX makes its proposed workflows and observed status understandable. Explore how source records could become a human-reviewed finding below: this is a synthetic guided example, not live AI inference or customer processing.
Concept exhibit · The future of business intelligence
A command room for evidence, not for guesswork.
Art direction inspired by contemporary Gulf architecture—vertical light, smoked glass, measured brass. Original artwork; no affiliation, endorsement, hosting location or live telemetry is implied. Every module below is synthetic, sample or developmental unless a check says otherwise.
OX WORKSTATION Overview
Explore the audit process
- Railway connectionUNKNOWN · HOLD
Read-only checks run on the website server. Your browser never contacts the service directly.
- API checksUNKNOWN · HOLD
Health, capabilities, evidence, budget and jobs — fresh for 90 s. Last check INVALID / FUTURE TIME.
- Audit exampleSample audit available
Find a duplicate invoice, compare ticket vs invoice vs rate, review the source — fictional data, in your browser.
- AI / NVIDIAResearch · not deployed
Adapters exist only as draft source branches. No website inference is enabled; model operation is not independently verified.
Development integration
Authored status · 2026-10-27- API source / testssource · SOURCE REPORTED · CI UNKNOWN
Historical branch-local tests and drafts are source evidence only. Exact deployed-head CI artifacts are not supplied; CI remains UNKNOWN / HOLD.
- Railway deploymentdeployed · DEPLOYMENT SUCCESS REPORTED
Owner supplied direct account inspection: staging October 24 and production October 22; owner reports staging aurex-api online October 27. Deployment success proves neither live endpoint health nor end-to-end operation.
- Staging Qwen inferencedeployed · OWNER-REPORTED · 2026-10-27
Owner reports qwen-private-staging deployed with Qwen2.5-0.5B-Instruct loaded; Railway logs show a processed prompt and an internal authenticated /v1/xaeli-chat/ask check returned text. Internal checks are not independent verification or general-chat proof.
- Website ↔ XAELI chatend-to-end · USER-OBSERVED WORKING · STAGING · 2026-10-27
User/owner-reported mobile website screenshot · 2026-10-27. User/owner reports a mobile website screenshot showing an open-ended audit-preparedness cost question and a contextual XAELI response. This supports a user-observed browser chat interaction in staging, not independently verified transport, inference, model provenance or answer accuracy. No cryptographic verification, generalized availability, audit approval, storage isolation or production-security proof; production readiness HOLD, financial authority NONE.
- Website ↔ backend authenticationend-to-end · UNKNOWN / HOLD
Only the current complete, fresh authenticated five-GET projection can establish read-only transport. Authored deployment facts never substitute for this check.
- Worker / SQLite ↔ websiteend-to-end · UNKNOWN / HOLD
A storage attachment is not persistence proof. Current worker operation, job/evidence read-back, restart durability and independent operational verification are UNKNOWN. Storage isolation and recovery remain blocked.
Current authenticated transport · last checked: NOT REPORTED / INVALID TIME · INVALID TIME · INVALID TIME
- staging · deployed 2026-10-24
- SUCCESS · one replica · storage attached
- Branch main · SHA f2db438f0020dbd455bcb5c7b21b7352d133bf2c
- production · deployed 2026-10-22
- SUCCESS · one replica · storage attached
- Branch integration/kusanagi-xvi-railway · SHA 57a079f93b947594f1a6f854aa46a6f606f86f76
Source: owner-supplied direct Railway account inspection, authored 2026-10-27; deployment dates have no supplied time-of-day. Scope: deployment records only, not website-observed health or live telemetry.
CRITICAL deployment blocker: storage isolation, full recovery and rollback NOT VERIFIED. Storage mount ≠ persistence. No writes or deployment authorized.
Independent operational verification UNKNOWN · staging browser chat user-observed, answers unverified · financial authority NONE · production release HOLD · paid external inference OFF. Evidence → Qualification → Reality → Authority → Effect. No self-witness.
Historical October 23 operator evidence · not current telemetry
Historical October 23 progress · operator report
Owner-verified Chromebook / Oct 23, 2026 · 2026-10-23 · not live telemetry
- NVIDIA model catalog
- CATALOG LISTED
- nvidia/nemotron-3.5-lightning-30b-a3b appeared in an authenticated catalog GET. Zero inference calls; entitlement and accuracy NOT VERIFIED.
- Railway staging access
- STAGING AUTH PASS
- Agent status returned authenticated HTTP 200 externally and inside the container. Running provider: OPENAI; chat configured: false. This does not prove NVIDIA deployment.
- NVIDIA backend configuration
- SAVED · NOT DEPLOYED
- Owner reports staging configuration saved. Inference and billing flags remain false; website-side staging availability NOT VERIFIED.
- SQLite local snapshot
- INTEGRITY CHECK PASSED
- The identified database snapshot passed its integrity check. Full volume recovery, storage isolation and rollback remain NOT VERIFIED.
- XAELI website model link · historical October 23
- NOT LIVE THEN · SUPERSEDED OCTOBER 27
- This October 23 observation predates the user-observed working staging website chat on October 27. The NVIDIA owner-only bridge remains a draft candidate; neither observation grants financial authority or independent model verification.
AUTHORITY NONE · RELEASE HOLD · Unknown ≠ verified. Catalog, transport and local integrity do not prove model accuracy or recovery readiness.
Evidence → Qualification → Reality → Authority → Effect. Current transport results stay separate; this report grants no execution or spending permission.
What is connected — and what remains · technical details
- Live transportCheck unavailable · HOLD
Only a fresh complete check establishes five authenticated read-only requests (health, capabilities, evidence, budgets, jobs). Deployment records never prove execution, inference or persistence.
- Lab tested · Oct 22, 202629/29 local checks passed
18 Python and 11 Worker tests on a development machine, plus a local database integrity check and a local 5/5 read-only bridge. These are local development results, not production proof.
- Operations not verifiedIn development
Deployment success and attached storage are historical facts. Worker execution, restart persistence and independent end-to-end verification remain UNKNOWN / HOLD.
Platform vs. API: the owner's Railway account can change infrastructure (deploy, configure). That is separate from what the running AUREX API does today, which these pages only inspect through read-only requests. Write access to the platform does not mean the product processes anything.
NEXT STEPS
- Safely preserve existing uncommitted work.
- Resolve storage isolation and recovery before any deployment or write.
- Prove database persistence across restarts, plus job lease and cancel.
- Deploy the authenticated worker and website only with owner approval.
- Monitor uptime around the clock and measure it.
- Enable a model only with explicit consent and an approved budget (today $0).
Missing or unknown evidence means HOLD. The system may never be its own sole witness; people keep final authority.
EXAMPLE WORKFLOW
Field tickets & invoices
- Input
- Fictional tickets, invoice IDs, approved rates and hours.
- Check
- Identify the records and the question being reviewed. Real customer intake is not available.
- Outcome
- A bounded example case with its source references.
Technical status & limitations
Only the existing read-only Railway service connection is checked. Last check: NOT REPORTED / INVALID TIME (INVALID / FUTURE TIME). A check expires after 90 seconds; missing, failed or stale checks never establish safety. No model inference is enabled; the inference budget is $0.
New audit, document and NVIDIA work is draft research, not deployed capability or independently established accuracy. Private jobs and evidence never enter this public page. The example steps above are authored explanations, not service records.
Evidence → Qualification → Reality → Authority → Effect. The system may never be its own sole witness. Unknown is not safe; financial authority stays HOLD for human review. Artwork motion grants no authority.
RESEARCH STAGE · NOT DEPLOYED
How NVIDIA research could help audits
A design map of how publicly documented NVIDIA reference blueprints could fit inside AUREX's evidence boundary. These references are research, not a NVIDIA partnership or running models.
NIM INTEGRATION MILESTONE
LOCAL CANDIDATE · NOT DEPLOYED
NVIDIA NIM adapter: prepared and tested locally
- Adapter code exists and was exercised against simulated API responses — no live model was called.
- Not deployed to Railway. The running service still performs no model inference.
- NVIDIA model connectivity is not established: it needs the owner's provider account and an explicit release approval.
- Connected today: the website's five read-only checks to Railway. A 24/7 agent and hosted inference are not live.
- 01
Source / request
A question, document or case enters with its stated scope.
- 02
Evidence + scope
Sources, provenance and limits are recorded before any reasoning.
- 03
NVIDIA technology research
Reference blueprints studied for evaluation, signals, embeddings, distillation and fraud scoring.
- 04
Challenge the interpretation
A separate challenger must test the result. Not built; no measured results.
- 05
Evidence + human decision
Independent evidence and a person decide. The system is never its own sole witness.
- 06
XAELI explanation
Plain-language summary of what was checked, unknown and withheld.
Research nodes (step 03)
Find quantitative patterns
NVIDIA Nemotron + NeMo Agent Toolkit
Study unusual patterns in rates or charges and test alternative explanations.
Official NVIDIA blueprintRead transaction sequences
RAPIDS / NeMo transaction embeddings
Explore how sequences of transactions could add context to a record review.
Official NVIDIA blueprintResearch smaller domain models
Teacher / student · data flywheel
Study teacher/student evaluation before any domain-model training is considered.
Official NVIDIA blueprintSpot possible fraud for review
GNN + XGBoost + Dynamo-Triton
Explore candidate risk signals that would still need source checks and human review.
Official NVIDIA blueprint
Document research & development status
Document research would read invoices and contracts with cited source passages, so a reviewer can trace a proposed finding. It is not available for customer documents.
Owner-reported draft work: document integration PR #106, NVIDIA references PR #107 and contract candidate PR #108 are unmerged and not deployed. There is no enabled GPU or model inference, no measured audit accuracy, and no financial approval. Source research is not runtime verification.