About this session

Judith is a personal AI agent that has run in daily production for a year, used by more than a dozen of my family and friends across three WhatsApp group chats, Telegram, and Discord. This talk walks through how it's built, in two parts. The first part is the engineering that makes one agent safe for many people to share: a multi-tenant permission model (read-only for my mom, exec for me), a memory stack — FAISS + Neo4j + curated long-term notes — that stays useful over a year instead of bloating into noise, cron-scheduled subagents that scout and act on their own, and the guardrails it enforces on every message — redact personal info before posting to a group, never reply to the wrong person, and screen attacker-controllable text for prompt injection before acting on it. The second part takes the agent off the screen and onto a $50 pair of smart glasses. It captures what I see, describes and stores it as a running visual memory, sets destination path on maps before I get onto car, finds and tells me which aisle in the store to go to first, etc. I cover the latency budget that keeps it conversational — on-device Whisper for speech, cloud reasoning, sub-one-second round trips — and the custom neural voice it speaks in rather than stock TTS, drawn from my speech-synthesis background. Both parts are shown live, including a candid look at the pieces that don't work yet. Audience takeaways: A multi-tenant architecture for a personal agent multiple people actually share A memory design that survives real long-term use (not just a vector store) A defensive checklist for any agent that ingests untrusted text A blueprint for an ambient, vision-aware wearable interface on commodity hardware, with a real latency budget

Topics

Speaker

Sai Krishna Rallabandi

Director, Data Science · Fidelity Investments