Singapore's Foreign Minister's Homemade AI Built Using Raspberry Pi

Amid the heated debates on artificial intelligence (AI) regulations among world leaders, Singapore’s Foreign Minister, Dr. Vivian Balakrishnan, has taken a far more concrete step. Rather than just discussing it, he has openly introduced a personal AI assistant that he designed and built himself to support his daily diplomatic activities. Balakrishnan describes his virtual assistant as a “second brain” for a diplomat. Its capabilities include answering technical questions, conducting topic research, drafting speeches, and providing daily briefings. “This system has become invaluable; I wouldn’t even dare to turn it off!” he wrote in a Facebook post. Balakrishnan is not just a politician jumping on the technology bandwagon. He is a former ophthalmologist who graduated from the National University of Singapore and received a Presidential Scholarship in medicine in 1980, making him someone accustomed to systematic, evidence-based thinking. The virtual assistant system is built on two open-source foundations. The first is NanoClaw, an independent AI assistant based on the Claude model by developer Gavriel Cohen, which runs locally on a Raspberry Pi and connects directly to various messaging applications such as WhatsApp, Telegram, Slack, and Discord. Secondly, Balakrishnan applies the “LLM Wiki” pattern proposed by former Tesla AI Director Andrej Karpathy, as a solution to the AI’s “amnesia” problem, where it often loses conversation context with each new session. The system works by absorbing all of Balakrishnan’s speech drafts, articles, and web clippings, then processing them into a structured knowledge graph. Every time it receives a question, the AI performs a semantic query and injects relevant facts into its response, making it smarter the more it is used. For Balakrishnan, AI is not just a policy debate topic. “Diplomats who learn to work with AI will have a significant advantage, and I believe that advantage starts now,” he concluded. The tool named Mnemon stores information in a SQLite database, which is then converted into wiki pages and can be accessed via the Obsidian app on macOS or iOS. The semantic search process runs entirely locally using the Ollama platform on a Raspberry Pi 5, while voice message processing is done directly on the device via whisper.cpp. Thus, no confidential conversations have ever touched external servers. Furthermore, each chat group has its own Docker container and local memory, ensuring that data between groups remains securely isolated.