Neosys AI Solutions
Back to Blog & Insights
IOT · AI ENGINEERING

AI at the Edge: Running Inference on Hardware You Don't Control

An NFC smart stand or a plant-floor tablet doesn't have a GPU and can't always reach the cloud. Here's how we decide what runs on-device, what stays server-side, and what fails gracefully when the connection drops.

Neosys Engineering Team

Full-Stack Core

Aug 21, 2026 6 min read

💡 Key Takeaways & Architecture Summary

  • Edge inference needs a defined fallback for when connectivity drops the AI feature silently breaking is not an acceptable failure mode on plant-floor hardware.
  • Lightweight classification and anomaly-detection models run comfortably on stand-level hardware; anything needing an LLM-scale model stays server-side with a queued retry.
  • A 24/7 AI assistant on a customer-facing smart stand needs a stricter latency budget than an internal plant-floor tool users abandon a slow kiosk in seconds.

Deciding What Runs Where

We split edge AI into two buckets: small, deterministic models (anomaly thresholds, basic classification) that run directly on the connected hardware, and anything requiring an LLM-scale model, which stays server-side behind a queue with a defined retry and offline-fallback behavior.

Designing for the Connection Drop

Physical touchpoints an NFC stand, a plant-floor tablet lose connectivity in ways a typical web app never has to plan for. Every AI feature we ship on that class of hardware has an explicit degraded mode: cached last-known state, a simple non-AI fallback, or a clear 'reconnecting' state, rather than a silent failure the customer discovers on their own.

Tags:
#Edge AI
#IoT
#On-Device Inference
#Smart Retail
#AI Integration
RELATED SERVICE
RELATED SOLUTION

Related Engineering Articles

DATA ENGINEERING · AI

From Dashboards to Predictions: Adding Forecasting to an Existing Data Pipeline

Your SPC or OEE dashboard already tells you what happened. Layering a forecasting model on the same pipeline tells you what's about to here's the architecture change that takes, and the one it doesn't.

WORKFLOW AUTOMATION · AI

Agentic Workflows: When to Let AI Make the Call

Most workflow automation is if-this-then-that. Agentic automation adds a step where an LLM makes a judgment call dispatch priority, anomaly severity, a drafted reply with a human able to review before it ships. Here's where that boundary should sit.

Let's scope your first AI workflow

A 30-minute call, a shortlist of automations, and an honest estimate of what they're worth.