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 project

Core 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.

CategoryCandidate product / technologySuitable scenariosKey limitationsTechnology originCommercial licenceData & deployment constraintsActually testedLast verified
AI ModelsCommercial closed-weight model APIs (e.g. GPT / Claude / Gemini families)General Q&A, summarisation, coding assistancePay-per-use pricing; outputs change with model updatesNon-Chinese vendorsCommercial API subscription, pay per useCalls go through services abroad; cross-border data transfer needs assessmentNot tested—
AI ModelsOpen-weight models (e.g. Qwen / DeepSeek / Llama families)Private deployment, domain fine-tuningRequires own compute and operations; licence varies by versionOpen-source community and vendorsApache-2.0 / MIT / community licence — check per versionCan deploy on-premises; data can stay in-houseNot tested2026-11-13
Agent FrameworksOpen-source orchestration frameworks (e.g. LangChain / LangGraph / LlamaIndex)RAG, tool calls, multi-step workflowsFast version churn, unstable APIs; some server runtimes carry a different licenceOpen-source communityMIT (core); check the server runtime separatelyCan deploy on-premisesNot tested2026-11-13
Agent FrameworksCloud-vendor managed agent servicesFast pilots, managed operationsVendor lock-in; limited customisationNon-Chinese cloud vendorsCommercial subscriptionData is processed by the cloud vendor; assess before useNot tested—
Cloud & ComputePublic-cloud GPU compute (on-demand / reserved instances)Training, batch inferenceCost varies with usage; GPU supply can queueNon-Chinese and Chinese cloud vendorsCommercial subscriptionData residency follows the chosen regionNot tested—
Cloud & ComputeOn-premises GPU serversContinuous inference, data stays in-houseHigh upfront cost; operations are your ownMulti-vendor hardwareHardware purchase + open-source softwareData stays in-houseNot tested—
Edge ComputingEdge AI compute modules (e.g. NVIDIA Jetson Orin family)Industrial vision, local inferenceCompute and power ceilings; fast model turnoverNon-Chinese vendor (NVIDIA)Hardware purchaseData processed locallyNot tested2026-11-13
Edge ComputingIndustrial edge gateways and IPCsEquipment networking, protocol conversionProtocol adaptation takes significant workMulti-vendor (wide choice of Chinese vendors)Hardware purchaseData processed locallyNot tested—
Vision & SensorsIndustrial cameras and sensorsDefect inspection, dimensional measurementSelection depends on site lighting and line speedMulti-vendor (wide choice of Chinese vendors)Hardware purchaseImage data processed locallyNot tested—
Vision & SensorsOpen-source vision inference frameworks and models (e.g. ONNX Runtime / OpenVINO / YOLO family)Object detection, classificationAccuracy depends on dataset and tuning; YOLO family is AGPL-3.0 — closed-source products need a commercial licenceOpen-source community and vendorsMIT / Apache-2.0 / AGPL-3.0 — check per projectCan deploy on-premisesNot tested2026-11-13
Industrial Systems IntegrationIndustrial protocols and middleware (OPC UA / MQTT / Modbus)Equipment networking, data acquisitionLegacy equipment speaks inconsistent protocolsInternational standards and open-source implementationsStandards are open; check each implementation's licenceData processed locallyNot tested2026-11-13
Industrial Systems IntegrationBusiness-system integration (ERP / MES / CRM, via API and MCP)Work orders, data write-backInterface documentation and permission governance take significant workClient's existing system vendorsUnder the client's existing licencesData handled within the client's system boundariesNot 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.