Think deeply · Deliver on site · Earn trust
Three layers of capability. Each answers the same questions: what it solves, how we do it, the evidence, and the products behind it.
Refining raw data into intelligence a model can consume directly
The problem
Enterprise data is fragmented and heterogeneous: spread across business systems, documents, spreadsheets and unwritten rules. Fed to a large model as is, it produces hallucinations and answers nobody can stand behind. What is missing is not volume but structure, lineage and shared definitions.
How we do it
- 1Compliant data pipeline
Raw / Clean / Curated layering; every cleansing and judgement rule is versioned, comparable and reversible.
- 2Dual-axis knowledge graph
Horizontal: global trade relationships up and down the value chain. Vertical: domain expertise. Every node carries source, evidence grade and validity range.
- 39Xin Compliant Acquisition Engine
An exception-hooking, multimodal in-memory engine that adapts to structural change and accesses publicly available business information within defined compliance boundaries.
- 4De-identification inside the client's domain
Raw data is de-identified and vectorised where it lives; only refined data reaches upper-layer computation.
Evidence
- 800M+ business entities across 200+ countries and regions (coverage figure); 3M+ curated high-quality trade records (refined figure)
- Vertical knowledge graph in production: 740K+ product records in the fact layer; 811 de-duplicated active ingredients × four-layer causal chain (vertical-industry scale, not a cross-industry figure)
- Text-generation algorithm for global enterprise data filed with the Cyberspace Administration of China: 网信算备440309441584901250013号
Related products
The model is an engine you can swap. The harness and skills are the chassis that accumulates value.
The problem
Models depreciate continuously. The right model today is replaced by a cheaper, stronger one within months. Capability tied to a single model has to be rebuilt every time.
How we do it
- 1Hybrid model routing
Routes by task type, data sensitivity and cost between local and cloud models. In private deployments, inference on sensitive data runs locally only.
- 2Harness
Tool calling, memory management, permission sandbox and evaluation loop. The 9Xin Iron Claw Security Engine is rebuilt at the Rust layer with 48 Skill and Harness classes.
- 3Versioned skill library
Repeatable expert workflows packaged as versioned skills, each with a golden set; upgrades must pass the evaluation gate first.
- 4Model-agnostic API
Applications call graphs and skills through one interface. Swap the model without losing capability; no single-vendor lock-in.
Evidence
- 9Xin Iron Claw Security Engine in production: Rust-layer rebuild, 48 Skill and Harness classes
- AI Maker Box ships with a sandbox runtime, security console and hybrid routing engine; live within days
- Sensory-science laboratory: model-swap SOP, evaluation gate and model-agnostic adapter delivered in phase one
Related products
We do not sell tools. We do not sell hours. We deliver three assets that grow together.
The problem
Tool vendors deliver a harness and leave the client to fill in context and skills; when they cannot, the project peaks on launch day. Consultancies deliver one-off skills and take them away; the next project starts from zero.
What we deliver is not a piece of software but a colleague who already knows the client's business.
How we do it
- 1Co-defined standards
We sit with the client's domain experts and write down what counts as right: judgement criteria, definitions, boundaries. The hardest and most valuable step in the project.
- 2Staged deployment
Each scenario is broken into stages, each with explicit inputs, outputs and acceptance deliverables. Prove it in one department, then replicate.
- 3Side-by-side coaching
Engineers work alongside frontline staff, and their corrections flow back into training and evaluation sets. Not a one-off training session, but working through the early errors together.
- 4Staged acceptance
Five metrics: voluntary frontline adoption, declining manual-review rate, marginal cost per task, whether anyone will sign off, and whether data flows back.
Evidence
- A provincial cross-border trade public-service platform: GoGlobal in production, lead agents covering six industries
- A sensory-science laboratory: AI-native data platform phase one, governance layer + harness + skill library, first skill packages delivered
- Replicated to further provinces without rebuilding the foundation
Related products
Why the three assets and FDE are inseparable
Only three things in an AI-native system accumulate and appreciate:
Context
Structured, traceable domain data and knowledge that AI can consume directly
Built by AI-Ready Data (ammunition)
Harness
Tool calling, memory, permission sandbox, evaluation loop. The model is a swappable engine; the harness is the chassis that accumulates
Built by 9Xin Iron Claw (security foundation)
Skills
Repeatable expert workflows packaged as versioned skill packages. The growth curve of the skill library is the learning curve of the system
Extracted on site by FDE
Context can be refined remotely. A harness can be delivered as a standard. Skills can only be extracted on site, because a skill is the organisation's own standard for judging whether something is right, and that standard has never been written down. It lives in the experience of senior staff, in the unwritten conventions of the business, and in the reviewer's remark that "something here is not right".
FDE delivers all three growing together, and every engagement folds the reusable part back into the framework. The next client is faster and cheaper.
We standardise how it is done. We customise what counts as right.
Where each asset lands and what the client gets
| Context | Harness | Skills | What the client gets | |
|---|---|---|---|---|
| Cross-border trade & going global | Global entity graph, multilingual entity alignment, compliance boundary definitions | Compliance audit gateway for cross-border data flows; source traceability | Target-customer rules, industry screening criteria, outreach scripts | Leads become opportunities you can act on; acquisition cost falls measurably |
| Smart manufacturing & inspection | Defect criteria, process parameters, Raw / Clean / Curated line data | Edge sandbox execution; permission isolation from line systems | Changeover judgement skills, inspection parameter recipes | New scenarios adapted within weeks; product changes do not mean a rebuild |
| Education | Structured teaching resources, longitudinal learning records | Private deployment on the campus network; multi-agent scheduling | Subject teaching-assistant skill packages | Data never leaves the campus network; affordable integrated hardware and software |
Compounding value, in a form you can write into a contract
Not a slogan, but four mechanisms:
- 1Versioned rule evolution
Every revision of a cleansing or judgement rule is versioned, reversible and comparable
- 2Bad-case feedback
Frontline corrections flow automatically into training and evaluation sets
- 3Evaluation sets
Every skill has a golden set; upgrades must pass the evaluation gate
- 4Data-model co-evolution
Service precision improves with usage
Five metrics that separate real AI projects from demos
- 1Voluntary frontline adoption
- 2Declining manual-review rate
- 3Marginal cost per task
- 4Whether anyone will sign off
- 5Whether data flows back
If launch day is the peak, it was a demo. If it launches imperfect and is better three months later, that is real productivity.
Start with one concrete scenario, not a product list.
Tell us which scenario is still not in production: where the data lives, who uses it, and who signs off. One workshop gives you a verifiable path forward.