TECHNOLOGY ECOSYSTEM
Technology Ecosystem & Selection
Negency helps enterprises choose across global AI technology and China's engineering and hardware supply chain: business problem first, then the technology combination, then verification in a pilot.
Assess my AI projectCore judgement. China’s supply-chain advantage is not a lowest-price promise — it is multi-vendor choice, hardware-software combinations, engineering iteration, and delivery flexibility.
Partnerships. The products and technologies on this page are selection research subjects; Negency has no confirmed formal partnership with any listed vendor. Naming a product only defines the selection scope; it is not a recommendation or endorsement.
Compliance. Cross-border data and regulatory requirements are assessed per project.
MODULE 01
AI Models
Technology options
Two routes: commercial closed-weight model APIs (fast to start, pay per use) and open-weight models (deployable on your own infrastructure, fine-tunable). One business problem often needs a mix — a small model for frequent routine tasks, a large model for complex judgement.
How Negency selects
First the constraints: can data leave the premises, what are the latency and cost ceilings, is fine-tuning needed. Then 2–3 candidates are compared on the client's own evaluation samples — not on public leaderboards alone. Finally the licence terms are checked to confirm commercial deployment is allowed.
What the client provides
The business problem and evaluation samples (questions with expected answers), data-export and compliance requirements, budget and latency ceilings, and a list of existing systems.
How it is verified
Candidates are compared offline on the client's own evaluation samples; during the pilot, accuracy, cost, and failure cases are recorded; before launch, permission and data-boundary checks are run.
Main risks
Model updates change outputs; commercial APIs raise data-export and vendor lock-in questions; open-weight licences vary by version and must be checked item by item before deployment.
MODULE 02
Agent Frameworks
Technology options
Open-source orchestration frameworks (LangChain / LangGraph, LlamaIndex and similar — mostly MIT-licensed at the core) and managed agent services from cloud vendors. A framework only solves orchestration; permissions, approvals, and audit have to be designed separately.
How Negency selects
By workflow complexity: lightweight setups for single-step retrieval, state-machine orchestration for multi-step flows with branches and human approvals. Core and server-runtime licences are checked separately — some projects license their production server runtime differently from the core.
What the client provides
The target business process and its exception rules, the systems to connect with their interface documentation, and the approvers and permission matrix.
How it is verified
The normal flow plus four exception classes are run in a controlled environment: no answer found, unauthorised tool call, tool failure, approval refused. Every step's audit record is checked for completeness.
Main risks
Frameworks iterate fast and APIs change; managed services lock you to one vendor; errors compound across multi-step agents, so human approval is the backstop.
MODULE 03
Cloud & Compute
Technology options
Public-cloud GPU compute (on-demand or reserved), on-premises GPU servers, or a mix of both. Training and batch inference suit elastic cloud compute; continuous inference and data that must stay in-house suit local deployment.
How Negency selects
By workload shape and data boundaries: intermittent heavy loads go to the cloud; 24/7 inference is costed locally over its full lifetime. Regional compute supply and data-residency options are checked for the target region.
What the client provides
Workload characteristics (peaks, continuity), data-residency and compliance requirements, existing IT infrastructure, and budget structure (capital vs operating expenditure preference).
How it is verified
Small-scale cost measurement first — the same task run under each option, recording actual spend and elapsed time — then the deployment shape is decided. Cost and utilisation are monitored continuously after launch.
Main risks
Tight GPU supply means cloud queues or price rises; on-premises servers need large upfront investment and self-run operations; cross-border data transfer needs a per-project compliance assessment.
MODULE 04
Edge Computing
Technology options
Edge AI compute modules (such as the NVIDIA Jetson Orin family), industrial edge gateways, and industrial PCs. The value of the edge is low latency, operation without connectivity, and data that never leaves the plant.
How Negency selects
By model compute needs, power budget, and site conditions; industrial sites favour wide temperature range, vibration resistance, and long supply lifecycles. The software stack (JetPack and similar) must match the model's inference framework.
What the client provides
Site conditions (temperature, vibration, power), model and latency requirements, existing equipment and network layout, and maintenance capability.
How it is verified
Offline verification in the lab with the real model and real data first, then a small on-site pilot; latency, power draw, and long-run stability are recorded.
Main risks
Edge compute has a ceiling — large models will not run; models turn over fast, so end-of-life risk is assessed at selection time; conditions beyond a module's rated range cause instability.
MODULE 05
Vision & Sensors
Technology options
Industrial cameras and sensors (multi-vendor, with a wide choice of Chinese vendors) and open-source vision models and inference frameworks (ONNX Runtime, OpenVINO, the YOLO family and similar). The camera decides input quality, the model decides judgement — the two are selected together.
How Negency selects
Cameras and lighting by inspection target, line speed, and site lighting; models and inference frameworks by accuracy requirements and licence terms — note that some popular models carry a strong copyleft licence (AGPL-3.0), so closed-source products must confirm a commercial licence first.
What the client provides
Defect definitions and sample images (a batch each of pass and fail), line speed and accuracy targets, site lighting and mounting conditions, and false-positive tolerance.
How it is verified
Offline verification on the client's real samples (accuracy and false-positive rate reported separately), then an on-site pilot; every image judged defective must be replayable with its evidence and review record.
Main risks
Accuracy figures are meaningless without enough samples; lighting changes discount lab accuracy on site; strong-copyleft models in closed-source products carry legal risk.
MODULE 06
Industrial Systems Integration
Technology options
Industrial protocols and middleware (OPC UA, MQTT, Modbus and similar) connect equipment; APIs and MCP connect business systems (ERP / MES / CRM). The hard part of integration is rarely the protocol itself — it is legacy equipment with inconsistent protocols and incomplete interface documentation.
How Negency selects
Open standard protocols first (OPC UA is the IEC 62541 open standard; MQTT is ISO/IEC 20922). Integration depth follows the client's existing systems: read-only data collection first; write-back operations require permissions and approval.
What the client provides
Equipment list and protocol documentation, business-system interface documentation and test accounts, data boundaries and permission requirements, and maintenance windows and acceptance criteria.
How it is verified
Integration is connected in a test environment first, then rolled out on site in stages; every class of write-back operation passes permission checks, human approval, and audit logging; acceptance is checked item by item against pre-agreed criteria.
Main risks
Legacy equipment without standard interfaces means large adaptation effort; business-system upgrades can break integrations; a wrong write-back affects production, so a rollback plan is required.
SELECTION MATRIX
Technology Selection Matrix
This page explains the selection method for six technology modules, with a checkable selection matrix — every row states plainly whether it has actually been tested. Tested means tested; untested is written as untested.
Every row in the matrix is checkable: “actually tested” means Negency has run a real test of the product; “last verified” is the date the row’s public facts (name and licence type) were last checked against public sources — it is not a test date.
| Category | Candidate product / technology | Suitable scenarios | Key limitations | Technology origin | Commercial licence | Data & deployment constraints | Actually tested | Last verified |
|---|---|---|---|---|---|---|---|---|
| AI Models | Commercial closed-weight model APIs (e.g. GPT / Claude / Gemini families) | General Q&A, summarisation, coding assistance | Pay-per-use pricing; outputs change with model updates | Non-Chinese vendors | Commercial API subscription, pay per use | Calls go through services abroad; cross-border data transfer needs assessment | Not tested | — |
| AI Models | Open-weight models (e.g. Qwen / DeepSeek / Llama families) | Private deployment, domain fine-tuning | Requires own compute and operations; licence varies by version | Open-source community and vendors | Apache-2.0 / MIT / community licence — check per version | Can deploy on-premises; data can stay in-house | Not tested | 2026-11-13 |
| Agent Frameworks | Open-source orchestration frameworks (e.g. LangChain / LangGraph / LlamaIndex) | RAG, tool calls, multi-step workflows | Fast version churn, unstable APIs; some server runtimes carry a different licence | Open-source community | MIT (core); check the server runtime separately | Can deploy on-premises | Not tested | 2026-11-13 |
| Agent Frameworks | Cloud-vendor managed agent services | Fast pilots, managed operations | Vendor lock-in; limited customisation | Non-Chinese cloud vendors | Commercial subscription | Data is processed by the cloud vendor; assess before use | Not tested | — |
| Cloud & Compute | Public-cloud GPU compute (on-demand / reserved instances) | Training, batch inference | Cost varies with usage; GPU supply can queue | Non-Chinese and Chinese cloud vendors | Commercial subscription | Data residency follows the chosen region | Not tested | — |
| Cloud & Compute | On-premises GPU servers | Continuous inference, data stays in-house | High upfront cost; operations are your own | Multi-vendor hardware | Hardware purchase + open-source software | Data stays in-house | Not tested | — |
| Edge Computing | Edge AI compute modules (e.g. NVIDIA Jetson Orin family) | Industrial vision, local inference | Compute and power ceilings; fast model turnover | Non-Chinese vendor (NVIDIA) | Hardware purchase | Data processed locally | Not tested | 2026-11-13 |
| Edge Computing | Industrial edge gateways and IPCs | Equipment networking, protocol conversion | Protocol adaptation takes significant work | Multi-vendor (wide choice of Chinese vendors) | Hardware purchase | Data processed locally | Not tested | — |
| Vision & Sensors | Industrial cameras and sensors | Defect inspection, dimensional measurement | Selection depends on site lighting and line speed | Multi-vendor (wide choice of Chinese vendors) | Hardware purchase | Image data processed locally | Not tested | — |
| Vision & Sensors | Open-source vision inference frameworks and models (e.g. ONNX Runtime / OpenVINO / YOLO family) | Object detection, classification | Accuracy depends on dataset and tuning; YOLO family is AGPL-3.0 — closed-source products need a commercial licence | Open-source community and vendors | MIT / Apache-2.0 / AGPL-3.0 — check per project | Can deploy on-premises | Not tested | 2026-11-13 |
| Industrial Systems Integration | Industrial protocols and middleware (OPC UA / MQTT / Modbus) | Equipment networking, data acquisition | Legacy equipment speaks inconsistent protocols | International standards and open-source implementations | Standards are open; check each implementation's licence | Data processed locally | Not tested | 2026-11-13 |
| Industrial Systems Integration | Business-system integration (ERP / MES / CRM, via API and MCP) | Work orders, data write-back | Interface documentation and permission governance take significant work | Client's existing system vendors | Under the client's existing licences | Data handled within the client's system boundaries | Not tested | — |
Rows marked “—” have not been individually checked and serve as selection references only. All 12 rows are currently untested; test status and dates will be updated here once pilots produce results.
NEXT STEP
Start selection from your scenario
Tell us the business problem and the constraints first; we will propose candidate combinations and a verification plan — no product is pushed before that.