CONNECTED TECHNOLOGY. CLEARLY EXPLAINED.
AI, Cloud & Edge
Choose device, edge and cloud processing through latency, uncertainty, data permissions, recovery and ongoing operating cost.
Put computation where the job can be performed and maintained reliably. AI adds an inference or prediction, but the complete system still needs trustworthy inputs, observable errors and bounded actions.
Choose where the job should run
On-device inference
Evaluate the actual model on the target hardware and representative inputs.
Edge gateways
Plan buffering, local rules and recovery when upstream services fail.
Cloud services and agents
Specify access, usage limits, retained data and actions that require approval.
Explore intelligence, infrastructure and evidence
Assistants and agents
Define the task, available tools and actions an assistant may perform. Keep observations and proposed actions traceable, set operating limits and name the approvals required for consequential steps. A fluent response does not establish that a device action succeeded.
Voice assistants · AI agents · Workflow agents · Multimodal assistants · Device control assistants · Agent orchestration.
On-device intelligence
Evaluate the actual model and runtime on the target device with representative inputs. Record memory use, processing delay, energy context and error behaviour; a chip capability or demonstration alone does not establish deployment performance.
TinyML · Embedded inference · NPU modules · Model compression · Speech processing · Offline vision.
Cloud IoT platforms
Review device identities, provisioning, message handling, update management and account roles as separate services. Document ongoing costs, retained records, export and the path to recover or move devices if a service changes.
Device registries · IoT message brokers · Device provisioning · Remote device management · OTA services · Fleet dashboards.
Edge computing
Place local processing around a defined latency, availability or data-handling need. Test buffering, duplicate handling, versioning and recovery across upstream interruptions; local execution still needs maintenance and observable failure states.
Edge gateways · Industrial edge servers · Edge orchestration · Container runtimes · Edge observability · Offline synchronization.
Analytics and digital twins
Distinguish measured inputs, calculated estimates and modelled scenarios. State the represented system, assumptions, update timing and validation method. A visually detailed twin is useful only when its outputs are appropriate for the intended decision.
Stream analytics · Time series databases · Sensor fusion · Anomaly detection · Predictive models · Digital twin platforms.
Vision audio and sensing AI
Match input quality and evaluation examples to the real environment. Review missed events, false alerts, uncertain outputs and affected users. Keep sensing permissions, retained media and any resulting physical action within the approved scope.
Computer vision · Acoustic intelligence · Predictive sensor models · Multisensor learning · Event classification · Synthetic data.
Data engineering
Preserve source identity, units, timestamps, quality flags and processing versions through the pipeline. Define how late, duplicated, missing or corrected observations appear so that a downstream chart does not hide uncertainty in the original record.
Data pipelines · Data lakes · Event streaming · Semantic layers · Data quality tools · Metadata catalogs.
AI governance and reliability
Assign model and service owners, evaluation criteria, change review and an incident response. Reassess behaviour when data, models, configuration or connected tools change. Make stopping or returning to a suitable existing process an explicit design decision.
Model evaluation · Agent permissions · Prompt injection defenses · Privacy-preserving AI · Human approval workflows · Model lifecycle management.
A practical example
For a proposed visual inspection assistant, reserve representative examples outside development data. Record missed defects, false alarms, processing delay and uncertain cases. Compare with the existing process before allowing automatic rejection or physical action.
Questions to resolve
- What happens when confidence is low or an input is stale?
- How are model and configuration versions recorded?
- Can the system stop or fall back when a service changes?
Use the evidence
The category map describes technologies and product types. It does not establish that a particular model is available, tested or compatible. Check the exact model, region, documentation and support arrangements before buying.
Work through the connected-technology buying checklist
Primary documentation and evaluation context
Google AI Edge: LiteRT for Microcontrollers supplies a constrained on-device runtime example. NIST AI RMF 1.0: core functions supplies voluntary risk-management context from the 2023 framework. Documentation describes intended capabilities and practices; it does not constitute an IoTPapa benchmark, certification or guarantee of a proposed system.
Evidence: Category guide. Reviewed 2026-10-18. By Admin.
Predictions, interfaces and dated outlooks
RELATED READING
AI and IoT: from a prediction to a bounded decision
Connect predictions to measured errors, review and bounded actions.
RELATED READING
Internet of Senses: practical interfaces and speculative claims
Separate current interface evidence from speculative sensory claims.
RELATED READING
IoT outlooks from 2024 and 2025: retain the questions, check the evidence
Read dated outlooks as an archive and recheck current claims.