02 / Case Study
02Chief of Staff Agent
A real personal AI — without the runaway bill.
Most personal AI tools either cost too much to leave on, or have no real memory. This one solves both at the architectural layer.
A Telegram-based personal-ops agent that runs the calendar, schedules tasks, and remembers ideas long-term — engineered so the cheap router handles the routine traffic and expensive calls only fire when they have to.
Problem
Useful personal AI hits a cost wall.
An AI that’s actually useful day-to-day needs to (a) be always-on, (b) remember things across conversations, and (c) cross-reference past ideas. Calling a frontier model for every text message — even small ones — makes that economically unworkable for an individual or household.
Most off-the-shelf personal assistants either ship without real memory, or assume enterprise-scale token budgets. Neither is right for a couple managing daily life.
Strategic Approach
Tier the model. Route by intent.
Every incoming Telegram message first hits a very low-cost router model that classifies intent. Task-shaped messages (create a reminder, add to calendar) are handled inline by the cheap model — no expensive call needed.
Memory-shaped messages (an idea, a thought, building on something said earlier) get routed to memory: the cheap model embeds the full content into a vector database for later retrieval. Retrieval uses the cheap model too, only escalating when a question genuinely needs more reasoning.
Result: a pseudo-RAG system that delivers real long-term memory and cross-referencing at a fraction of the cost of naïve always-frontier calls.
How It Works
Telegram in. Calendar, tasks, and memory out.
Built with Claude Code on top of Telegram. The agent reads, edits, and moves calendar events. It creates tasks with scheduled reminders that push to phone notifications when they’re due. It stores ideas and thoughts long-term and can surface connections between them when asked.
The architectural trick — routing first, embedding second, escalating only when needed — is the reason it’s actually usable as a daily tool.
Demo
See it in motion.
Simulated, zero-token — feels live, costs nothing, can’t be abused.
Loading demo…
Outcomes
What it changed.
- Routing intent before generating keeps an always-on household agent affordable: the low-cost router classifies every incoming message, handles task-shaped ones inline, and only escalates to bigger models when a question genuinely needs them.
- The pseudo-RAG memory layer gives it something most cheap assistants don’t have — long-term recall of ideas and thoughts, with cross-references and connections between them.
- Day-to-day ops run from a Telegram thread the household already lives in: calendar reads, edits and moves, plus tasks with scheduled reminders and push notifications.