Artux
About Artux

Three systems, twelve months, all built in-house

Artux was incubated by Henglin Hengjun, the company behind Nouhaus, a cross-border DTC furniture brand. Starting in August 2025 we spent twelve months building three systems ourselves: the digital-employee runtime, the cloud control plane, and the training system. We did not build a product and then look for a use case — it grew inside our own cross-border e-commerce operation first: 14 digital employees on staff, 59 skill packs and 44 unattended scheduled jobs running in production every day, and only then was it productised. "Your data never leaves" is not a promise but a consequence of owning the whole stack: we wrote the runtime and the control plane, instances run on our own machines, and no third-party agent platform sits in the middle.

Scale of the in-house stack

production code across five in-house systems

1.23M lines

production code across five in-house systems

● code count

commits in twelve months

1,593

commits in twelve months

● code count

digital employees on staff · skill packs

14 · 59

digital employees on staff · skill packs

● code count

unattended scheduled jobs

44

unattended scheduled jobs

● code count

As of 2026-07-27, from repository statistics and production snapshots. "Code count" excludes tests, dependencies and build output. Demo and load-test figures are excluded and are never restated as business volume.

Twelve months, four phases

Every milestone below is a first-commit date you could check in git. We started where most cross-border companies start: we bought the AI tools on the market and used them. The problem was not that they were not smart enough — each covered one segment, none talked to the others, and none understood the specific rules of this business chain.

  1. 2025-08

    Where in-house work began

    First self-owned commit to the operations platform. Fix what hurts: product data, logistics tracking, the first plugins.

  2. 2025-12

    The first real AI capability

    The AI image platform kicked off, and the first in-house Shopify app shipped the same day. The tool phase ended; AI started entering the workflow.

  3. 2026-02

    Business becomes systems

    PIM sell-through and inventory plus the SpuKit spec app went live the same day; the ops and security practice was established across seven managed servers.

  4. 2026-04

    The turning point — artux-agent

    AI went from tool to colleague: 14 roles, 6 departments, 59 skill packs and 22.5k lines of role handbooks. Assigning work is one sentence in Feishu.

  5. 2026-06

    The GEO workstation

    Buyers began asking AI which product to buy. Day one produced the skeleton, nine connectors and the audit rule engine to hold that new entry point.

  6. 2026-07

    The gateway reaches the back office

    act-as-user mints a restricted credential as the requester themselves and passes through the existing permission system — a digital employee can never read what that person cannot.

  7. 2026-08

    Productised

    artux.ai opened for free sign-up. Same runtime: one digital employee equals one dedicated cloud instance, provisioned on request.

Three systems we built

What a customer sees is "a new colleague in Feishu". Three separate systems, with no overlapping responsibility, hold that sentence up.

01

Digital-employee runtime

what runs on the customer machine

A leader agent takes the work, routes it to specialist sub-agents and tracks it to completion. The desktop build and the cloud image share one kernel — nothing works in the cloud but not locally.

  • Feishu-native collaboration: messaging, Base, Docs, Tasks, Calendar, Attendance
  • Real browser control: CDP-driven Chromium, persistent login state, human takeover for CAPTCHAs
  • Long-term memory that survives restarts and staff handover; corrections apply immediately, stale facts expire
  • Logs a commitment when externally blocked, redoes the work when conditions are met and reports back honestly

02

Cloud control plane

provisioning, brokering, accounting

Tenancy and subscriptions, automatic provisioning of dedicated instances, credential brokering, usage and audit. The customer never touches a server or the code.

  • One company, one instance; physical data isolation, hardened containers, full audit trail
  • Zero credential landing: AI and app secrets stay in the cloud gateway, revocable per instance
  • Token usage, seats and audit events visible from the same source as the console — no silent overage
  • Instances report health and version back; the customer decides when to upgrade

03

Training system

how company SOPs are taught to an AI

A digital employee can do real work not because the model is stronger, but because somebody wrote down the rules of this chain and taught them. The methodology itself is never shipped, but it decides whether the delivered employee is any good.

  • 22.5k lines of role handbooks · 59 skill packs · 81 knowledge-base documents
  • Use it once, keep it: recurring workflows are proposed as standard skills and shared once confirmed
  • Behaviour changes are verified end to end from an operator’s seat, never by unit tests alone

Why build it ourselves

These are not values on a wall. Each one is a judgement we have paid for.

1

You can only teach a chain you have walked

Plenty of companies build agents, plenty build SEO tools. But teaching a full DTC chain to a digital employee assumes your own company has actually run it: where sourcing stalls, what a wrong replenishment number ties up in cash, how much conversion one bad line of theme CSS costs, what a carding bot on your checkout page feels like. That cannot be learned from documents, only from the holes you have fallen into.

2

Constraints in code are what make AI safe to bring inside

The answer to "will the AI go rogue" is not an assurance but a gate on the execution path: publish and write-back are not automatable by task type and are rejected at registration; after editing a storefront the agent may only push a preview with dual-viewport screenshots and must wait for an explicit "go live"; if an official API is unavailable the field is marked "not connected" and never fabricated.

3

Owning the stack and the infrastructure is what keeps data in

Runtime, control plane and training system are all ours, and instances run on our own AWS account and our own machine room. With no third-party agent platform in the middle, there is no question of where the data went. Customers manage everything through artux.ai and touch neither the code nor the instance host.

A digital employee grown inside a real business

Signing up is free and needs no credit card. If you want the boundaries first, the capability list marks every item stable, limited or roadmap.

Sign up freeSee the capability list
About Artux — three systems built in-house in 12 months