I rarely see anyone using AI models this way in forums, so I hope more people will see and use the features in these directions. 
The buzz around AI, particularly powerful models like Gemini, often centers on software or content generation. However, its potential to fundamentally reshape complex, tangible fields like electronic hardware development is immense and, I believe, underexplored. I’m here to share some insights and a real-world example of how AI can streamline and elevate the entire product lifecycle, from initial R&D to full-scale production.
### The Old Hurdles and AI’s New Paths
In the hardware industry, engineers often carry a wealth of experience. But sometimes, this experience, especially if it’s “expired knowledge,” can lead to path dependency and cognitive ruts. The sheer depth and breadth of AI’s knowledge often surpass that of most individuals, even without RAG (Retrieval Augmented Generation) enhancing its understanding of specific domains.
We’re seeing that AI’s problem-solving “thought processes” aren’t always aligned with human cognitive structures. This divergence can lead to surprisingly powerful optimizations or even “emergent” solutions that human designers might not have conceived. It’s a bit like how my Google API KEY can get exhausted; human cognitive and knowledge discovery processes also hit limits, especially when compared to an AI that never sleeps (even beyond a “996” work schedule!).
AI in Industrial Hardware
When new technologies emerge, they often first permeate sectors less sensitive to initial costs and longer timelines. That’s why I’m particularly bullish on AI’s role in developing industrial hardware . This contrasts with applications like visual production line inspection or highly personalized consumer products, where the current “tokenomics” (the cost of AI processing) can be prohibitive for high-volume, low-margin scenarios.
Today’s AI can effectively assist in creating sophisticated tools, like the visual inspection systems developed by companies such as Halcomm. However, it’s not yet feasible to, say, connect a laptop camera directly to an AI for real-time, video-stream-based quality control on a massive production line due to token consumption.
While some AI functionalities are best handled by tech giants like Google or Microsoft (think Jeff Dean’s foundational search algorithms), these will become crucial building blocks for widespread AI application across industries. I see current RAG implementations as a transitional phase, an attempt by many to reconstruct Google’s search prowess. Personally, I’d rather lean on Google’s own mature products.
The core impediment for many AI products with real-world integration remains the staggering consumption of tokens.
A Concrete Example: AI-Driven R&D for an Industrial Lighting Source
Let’s dive into a practical case: developing a spectrum-precisely-tunable professional industrial lighting source.
(An important aside: Google’s robust multilingual capabilities across its products are a significant competitive advantage. This is often undervalued by VCs but is critical in actual production environments. For instance, our documentation typically needs to be perfectly consistent across Chinese, English, and Thai.)
//Note that this generates SVG diagrams in both Chinese and English, rather than simply replicating online translation. The knowledge-based generation of these different language versions ensures accurate, industry-specific content, avoiding the translation errors commonly found in engineering terminology when using translation software.
If we enhance this process with RAG, using real product manuals from suppliers and cutting-edge academic papers, the accuracy of these SVGs can be further improved.
//There’s also an advanced usage: combining A2A or MCP tools to directly call project management web pages or tools for automation.
Next Steps: Build and develop a demo in the lab.
//In various module and circuit designs, we’ll use other AI tools and traditional EDA routing tools. AI tools can significantly reduce manual workload, and AI can verify the correctness and reliability of designs based on physics and electronics knowledge.
//Test procedures and plans written with AI can be used for real-world step-by-step verification. Note that some manual intervention is needed as current AI capabilities are insufficient and can make basic errors.
//The code in the MCU is mainly generated and tested by AI based on defined functions.
//Here’s a tip: accurately specifying algorithm processes or requiring the use of algorithms like Extended Kalman Filtering in the prompt can significantly improve code quality.
//When you need to perform long, tedious tests (such as temperature limit testing or aging tests), AI + camera is a good choice. Google’s models are the best in this area, especially in accurately recognizing abnormal fluctuations in oscilloscopes and actual products based on image recognition, far surpassing other models on the market.
There’s still a lot of room for AI applications in this industry. For example, the industrial sector actually needs specialized, miniaturized production line models, such as specialized inspection models controlled by larger “pro” models, trained only on production line content. This can greatly improve model response speed and reduce token consumption. Alternatively, these models can be deployed on local computing platforms and connected to cameras for data acquisition. (This is a very important consideration for many companies, as large production lines have data localization requirements and concerns about network outages or offline large models.)
If you have industry ideas or restructuring suggestions, please leave a comment below. I’m currently testing many industry rebuilds.
I also hope that Google’s next model training will give more consideration to real-world R&D and product build knowledge content and knowledge base construction. This can greatly reduce the deployment cost of RAG (because embedding knowledge directly in the model is much more efficient than manually enhancing the model with such knowledge).