Sage is my evolving personal AI system and working partner — not a chatbot. She began on Telegram; today she runs as native iPhone and Mac apps I built, on a Python multi-agent backend.
She runs like a CEO: I speak to Sage, and she routes each request to the right specialist — research, career, mail, operations — then brings the result back. Each agent has defined boundaries; Sage coordinates rather than doing everything in one prompt. I'm the product owner and final decision-maker.
● In the native apps
- Voice and text. Conversation works now; fluid interruption while she's speaking is still in development.
- Career support today — a broader local-first counsellor is in development.
- Limited inbox access today; fuller mail triage and drafting are still in development.
- Memory that carries context across sessions.
◆ Search & research
- Quick web search. A Tavily-powered pass fetches current information; Sage returns a short, cited answer in the conversation. If synthesis fails she hands back real source links rather than inventing one; if search is down, she says so.
- Deep Research. A dedicated agent plans focused sub-queries, searches several paths in parallel, weighs conflicting sources and re-plans once if a pass is thin — then returns a cited synthesis. It runs as a queued background job (minutes); the report lands in the app's Jobs list, with no completion alert yet.
◇ Built for honesty
Sage keeps an evidence-backed capability registry — every capability logged with its real dependencies, limits and status. A structural honesty tripwire catches routing, implementation and those registered claims drifting apart, so unsupported behaviour fails visibly instead of being passed off as working. Architectural honesty about what the system can do — not a promise that every answer is right.
Still evolving. Security-reviewed and continually hardened, with explicit boundaries and fail-visible behaviour by design; some findings remain open, including a production-auth foundation the career work surfaced. The backend is deployed; the native apps are founder-operated builds, not public consumer apps.
Not a bot — a personal AI system I designed, deployed, security-reviewed and keep hardening.
