Quest: Mandy - The Architecture of Privacy

Date: March 2026 Subject: High-Performance Agentic RAG for High-Sensitivity Domains

The Challenge: Building a Private AI Parenting Platform

Parenting data is among the most sensitive information a user can share. When building the Mandy Project, an AI-powered parenting platform, the primary goal was to provide high-level intelligence and guidance without ever compromising family privacy.

The Strategy: Architecture of Scarcity & Trust

Operating on self-hosted GPU hardware and using only Small LLMs (2B-9B), I focused on moving the "Intelligence" from the model weights into the Orchestration Layer, ensuring the engine is both powerful and transferable.

1. Privacy-by-Design Protocol

To protect family and developmental data, Mandy uses a stateless backend design. Personal context is managed by the client application — the backend processes only the immediate task and never retains sensitive user history. Privacy is enforced at the architecture level, not just by policy.

2. High-Reliability Orchestration

Giving parenting advice requires extreme reliability. I developed a custom orchestration framework that manages the flow of agents, allowing for:

• Resilient Execution: Automated retries and fallback logic for complex reasoning tasks.
• Multi-Tier Verification: A higher-tier model reviews the output of smaller models, ensuring high-quality, grounded responses.

3. Optimized Local Serving

Using optimized local model serving, I run multiple models on a single GPU — Embedding, Chat, and Reasoning — ensuring the platform is fast, private, and cost-effective.

The Outcome: The Transferable Mandy Engine

Mandy is a working proof that you can build intelligence into high-stakes domains like parenting using localized, small models. The underlying Mandy Engine is a decoupled, transferable RAG core that can be adapted for any domain requiring extreme privacy and high reliability.

Looking ahead: The platform runs on private hardware today. Private cloud inference — providers who guarantee no data retention or training on user data — is a clear path to reliability at scale, and it's on the roadmap.

---

High-Level Tech Stack: Python (FastAPI), Docker, React Native.