Nouhaus is a cross-border direct-to-consumer furniture company. Beginning in August 2025, we spent 12 months turning AI from a collection of purchased tools into an operating organization. By July 27, 2026, five in-house systems contained 1,227,041 lines of production code and 1,593 commits; 14 digital employees worked across six departments with 59 skill packages and 44 unattended scheduled jobs. Artux is the productized form of that experience. These are measured production or repository figures unless a passage explicitly says it is an estimate.
Why isolated AI tools were not enough
The tools were capable, but each owned only one fragment. A product research conclusion did not flow into supply planning. Generated imagery did not enter the listing workflow. An SEO report did not modify or verify the site. Every boundary created another manual handoff.
They also lacked the operating rules of a furniture business: heavy inventory, slow turns, seven warehouses, several countries, strict marketplace specifications, and production themes where one careless CSS change can hurt conversion.
Most importantly, after a year of individual adoption, the company had retained very little. Skilled users became faster, but their methods remained in personal accounts. That was the reason to build.
Four stages: tool, system, organization, connection
| Stage | Period | What changed |
|---|---|---|
| Tools | Aug–Dec 2025 | Product data, logistics tracking, the first AI image workflow, and the first in-house Shopify app |
| Systems | Jan–Apr 2026 | Profit modeling, design studio, PIM inventory analytics, international feeds, and operations security |
| Organization | Apr–Jun 2026 | 14 digital roles across six departments, 59 skill packages, and 22,000 lines of role manuals |
| Connected growth | Jun 2026 onward | GEO workstation, product collaboration, and the digital-employee gateway linking the operating platform |
The inflection point was April 20, 2026, when the digital employee system began. Earlier systems helped people calculate. The next system had to run work while nobody watched it. That change required a different level of identity, safeguards, recovery, and audit.
One product crosses five departments
A DTC furniture product moves from factory sample to product analysis, merchandising decision, creative production, listing, and performance monitoring. Waiting at those boundaries consumes margin.
The digital employee system turns the chain into explicit work:
- Sample data is entered once and reused downstream.
- Completed product analysis is proactively delivered to the responsible marketing owner.
- Approval creates purchase and photography tasks.
- Finished assets enter the listing workflow.
- A new listing enters inventory and GEO monitoring.
- Multi-day commitments are recorded, reminded, and reported instead of disappearing in chat.
By July 27, 2026, this workflow had produced 1,193 task artifact directories, 685 deliverable files, and 108 MB of deliverables, including reports, cleaned datasets, imagery, translations, and inspection results.
What changed across eight operating stages
Product selection
A single request can initiate pricing distribution, rating analysis, concentration metrics, opportunity mapping, and a report delivered to Feishu. The workflow itself has seven explicit states and alerts when a stage remains unchanged for three days or more. AI performs analysis; a deterministic system owns accountability.
Supply chain
Sales velocity over 7, 14, 30, and 90 days, available days, sell-through, and inventory health run at 07:30 every day. Inventory from Lingxing ERP, Shopify, Wayfair, Walmart Fulfillment Services, and Coupang synchronizes every four hours across seven warehouses in five countries.
Creative production
Ten specialized agents coordinate six stages. The system contains 41 furniture-oriented generation recipes, eight editing tools, and 15 selectable models. Every call records trace, purpose, token use, latency, error class, and USD cost to six decimal places. AI spend must be observable before image generation can become a production process.
Storefront and listing
Ten in-house Shopify apps total 259,000 lines. Four theme projects contain 1,184 Liquid files and roughly 291,000 lines. A digital ecommerce engineer can prepare changes, but it must inspect 1440×900 desktop and 390×844 mobile screenshots, provide a preview link, and wait for an explicit human approval before publishing live.
Technical SEO
The audit engine runs 29 rules covering technical, structural, content, GEO, and site-wide conditions. It connects search terms to specific Title, H1, body, and FAQ changes. It also calculates internal PageRank with a 0.85 damping factor over 50 iterations. In production it manages 122 pages, 9,907 internal links, 909 keywords, and 3,226 audit findings.
GEO
Generative Engine Optimization asks whether a brand can appear accurately in AI-generated answers. We trust official citation metadata rather than scraping answer interfaces; a provider without official citation data is marked “not integrated.” One complete AI citation-monitoring round costs $0.08 in measured production usage.
Finance
This stage intentionally avoids generative AI. A 44-account standardized P&L, five-dimensional aggregation, scenario analysis, and roughly 15 cost modules must remain reproducible and auditable. Deterministic work belongs to deterministic software.
Operations and security
Seven servers are managed over outbound SSH without installing agents on those servers. Security inspection includes 17 checks, custom WAF rules, malicious ASN controls, legitimate bot allowlists, and approval gates for repair commands. Backup covers five asset classes, including 203 GB of object storage, with two remote destinations and retention policies.
Constraints, not models, make unattended work possible
Our scheduler accepts only collection, analysis, and inspection jobs. It deliberately has no “write to production” job type, and tests enforce that absence. Production writes follow six rules: idempotency, explicit confirmation, a pre-write snapshot, rollback, complete audit, and a dead-letter path.
These controls are why the digital employee system itself reached about 90,000 lines, 224 test files, 318 database migrations, and 503 commits. Those figures describe the digital employee system, not the 1.23-million-line total across all five systems.
What became Artux
Artux packages the organizational layer: a dedicated company instance, controlled credentials, Feishu/Lark work assignment, role knowledge, health management, quota accounting, upgrades, backup, and recovery. The capability matrix shows what is stable, limited, unsupported, or planned; pricing states the current plan limits.
FAQ
How were these figures measured?
Repository figures exclude dependencies and build output. Production figures come from system and development-database snapshots in July 2026. Human effort comparisons are estimates and are labeled separately from measured system performance.
Does this apply outside furniture?
The eight-stage chain applies broadly to DTC operations. The parameters do not: furniture inventory horizons, warehouse behavior, and image recipes cannot simply be copied into fast-moving consumer goods.
Must a company build everything itself?
No. A tool stack can work, but it creates manual transfers between islands. Building becomes attractive when the same transfer happens every day and involves more than one person. Artux exists so another company does not have to repeat the full infrastructure journey.
Can AI accidentally publish a broken storefront?
The execution path prevents autonomous live publishing. It can prepare a preview and screenshots, but a human must explicitly approve the live step. Production writes also require idempotency, snapshots, rollback, audit, and a failure queue.
Should SEO and GEO be separate programs?
They share crawlability, structured data, page quality, and factual sources. GEO additionally requires answer-ready passages: an early conclusion, concrete evidence, clear entities, citations, and conversational question coverage.