Descriptions:
Steve Korshakov, co-founder of Bee AI (acquired by Amazon roughly eight months prior), describes the privacy-preserving architecture behind the company’s AI wearable at the AI Engineer conference. The device — worn on the wrist — continuously records audio and builds a persistent personal agent, capturing approximately 10 million tokens per person per year. Given the sensitivity of always-on audio (Korshakov notes a single week of recording reveals “virtually everything” about a person), the team designed the system so that neither Bee engineers nor Amazon employees can access user data.
The architecture rests on four principles: encryption keys stored exclusively on the user’s phone with no server-side persistence; mandatory end-to-end encryption with no opt-out path; public workload transparency via Sigstore (allowing any external auditor to verify deployed code); and confidential compute for all inference, including self-hosted model infrastructure. Key expiration is set at seven days — chosen because 24 hours proved too short for realistic usage patterns.
A two-tier signing system prevents unauthorized deployments: an independent privacy team manages the transparency log, and Bee’s own engineers cannot influence that team’s signing keys, which are hardcoded into client applications. Korshakov also frames the product’s stateful, always-running agent runtime as a preview of where personal AI is heading broadly — citing Claude Code’s evolution from request-response to multi-hour autonomous sessions as an analogous trajectory.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Deep Dive
