Exploring the Potential of TinyML at the Edge When the Cloud Isn’t Close Enough
Artificial intelligence is increasingly moving beyond the cloud and into the physical world.
From manufacturing equipment and connected infrastructure to wearables and consumer devices, billions of sensors continuously generate data. Traditionally, much of that data is transmitted to the cloud for processing and analysis. But as products become smarter and more connected, an important question emerges.
Does every decision need to travel to the cloud?
What happens when connectivity is unreliable, bandwidth is limited or a device needs to respond in real time?
This is where TinyML comes in.
At Actalent, we’re exploring how emerging technologies can be applied across the product engineering life cycle. Our recent TinyML Proof of Concept (PoC) brought together expertise in embedded systems, software engineering and machine learning to investigate how lightweight AI models can bring intelligence directly to resource-constrained devices.
Bringing AI Closer to the Data
Many connected systems still follow a familiar architecture:
Sensor → Cloud → Analysis → Decision
Cloud computing remains a critical part of the AI ecosystem, but it may not be the most practical solution for every application. Factors such as connectivity, data transmission and processing latency can become important considerations, particularly in distributed environments.
Our engineering team set out to answer a simple question: Can machine learning models perform analysis directly on a small embedded device?
To explore this, our engineering team developed a TinyML PoC using an ESP8266 microcontroller and a DHT22 temperature sensor. The goal was to demonstrate how sensor data could be analyzed locally, enabling the device to generate insights without relying on continuous cloud connectivity.
From Sensor Data to Local Intelligence
The PoC used temperature data collected from the DHT22 sensor and applied lightweight machine learning models developed using Edge Impulse.
Rather than relying solely on predefined thresholds, we explored how machine learning could enable the device to interpret sensor data in different ways.
Classification
The classification model identifies temperature conditions such as:
- Normal
- High
This demonstrates how a lightweight ML model can interpret incoming sensor data and categorize operating conditions directly on the device.
Regression
The regression model predicts the temperature value based on the data.
Together, these approaches demonstrate how TinyML can extend embedded systems beyond traditional rule-based monitoring toward more intelligent, data-driven decision-making. All inference takes place locally on the ESP8266 microcontroller. No continuous cloud connection is required.
While the use case itself focused on temperature monitoring, the broader objective was to explore how intelligence can be designed directly into edge devices.
Why TinyML Matters
As connected products generate increasing amounts of data, not every application requires that data to travel elsewhere before an insight can be generated.
For certain use cases, intelligence closer to the source could change the architecture from:
Collect data → Send to cloud → Wait for analysis → Receive response
To:
Collect data → Analyze locally → Generate insight
This shift could be particularly valuable in environments where responsiveness, connectivity or system resources need to be carefully considered.
For engineering teams, it also creates opportunities to design intelligent products that integrate hardware, embedded software and AI as a unified system rather than separate components.
The Engineering Challenge
Running machine learning models on powerful cloud infrastructure is one thing. However, running them on a resource-constrained microcontroller is another.
The ESP8266 has limitations in terms of memory and processing capability. This means deploying ML at the edge requires careful engineering across the development life cycle.
During the PoC, our team explored challenges including:
- Collecting and preparing meaningful sensor datasets
- Creating datasets suitable for both classification and regression
- Developing lightweight models
- Achieving reliable model performance
- Deploying models within embedded memory and processing constraints
This is where TinyML engineering becomes more than simply training a machine learning model.
It requires a combination of:
Embedded engineering + Sensor integration + Data engineering + Machine learning + Model optimization
Each element influences the other. The hardware determines available resources. The sensor data influences the model. The model must ultimately perform within the constraints of the embedded system.
This systems-level approach is critical when moving from an idea to a deployable intelligent product.
At Actalent, this convergence of disciplines is central to how we explore complex engineering challenges—bringing together capabilities across embedded systems, systems and software engineering and intelligent technologies to evaluate what is possible.
The goal is not necessarily to build the largest or most complex model, it’s to build the right model for the right device and application.
Beyond Temperature Monitoring
While this PoC focused on temperature data, the potential applications of TinyML extend across manufacturing, smart infrastructure, healthcare, agriculture and consumer electronics. Wherever devices generate data and need to make smarter decisions, local AI has the potential to improve responsiveness, reduce reliance on connectivity and enable more efficient data processing.
This exploration raises an important question for organizations developing connected devices: where could intelligence at the edge create value? While cloud-based AI will remain critical for many applications, others may benefit from hybrid or edge-first approaches.
Building Intelligence Where it Matters
The future of AI will not exist only inside large cloud infrastructure. Increasingly, intelligence will also exist at the edge—inside devices, machines, sensors and products interacting with the physical world.
Our TinyML PoC represents a small but meaningful exploration of that direction. More importantly, it reflects the growing need for engineering capabilities that can connect hardware, embedded software, data and AI into integrated solutions.
At Actalent, we bring together expertise across embedded systems, systems and software engineering, AI/ML and connected product development to help organizations explore, validate and engineer solutions for complex technology challenges.
Whether the starting point is an emerging technology, an engineering challenge or an early-stage product idea, our teams can help evaluate possibilities and translate them into practical solutions—from concept and Proof of Concept to product engineering and beyond.
Because sometimes the smartest place to process data isn't in the cloud, it's where the data is created.
