Project Mirror

Demo

Short walkthrough demonstrating multi-agent coordination, task execution, and real workflow automation.

---

Overview

Project Mirror is a multi-agent AI system designed to act as a professional assistant—handling tasks such as information retrieval, scheduling, and technical reasoning through coordinated agent workflows.

It is built to explore how AI systems can move beyond isolated capabilities into reliable, task-oriented tools used in real workflows, where outputs directly influence decisions and actions.

---

What This Solves

Most AI systems work well in isolation (chat, retrieval, automation), but struggle when combined into reliable, end-to-end workflows.

Project Mirror addresses this by:

• Coordinating multiple specialized agents to handle complex, multi-step tasks
• Integrating with external tools (e.g., Google Calendar, Google Meet) to perform real actions
• Enforcing structured outputs and validation to improve reliability
• Providing visibility into system behavior and failure modes

---

System Architecture

Project Mirror uses a hierarchical multi-agent architecture with a lean 3-agent core, designed to balance modularity with latency.

Core Agents

• Nexus (Orchestrator)
  Handles intent parsing, task decomposition, and execution.
  Also directly performs certain tasks (e.g., scheduling, technical reasoning) to reduce unnecessary delegation.

• Researcher (RAG Agent)
  Retrieves and grounds responses using Weaviate and external search APIs, with strict context isolation.

• Demo Specialist
  Handles sandboxed scenarios (e.g., customer support, data analysis) using controlled datasets.

---

Key Design Decision: “Zero-Hop Execution”

Earlier versions used deeper delegation chains across multiple agents.

The current system introduces a “zero-hop” execution model, where:

• The orchestrator handles certain tasks directly
• Delegation is used only when necessary

This reduces latency, cost, and error propagation while maintaining modularity where it matters.

---

graph TD
    User([User]) <--> Frontend[Next.js 16 / Tailwind CSS 4]
    Frontend <--> API[FastAPI Backend]

    subgraph "MAS Orchestration (Google ADK)"
        API <--> Nexus{Nexus Orchestrator<br/>Gemini 3.1 Flash-lite}
        Nexus -- "Control Handoff" --> DemoSpec[Demo Specialist<br/>Llama 3.3 70B]
        Nexus -- "Tool Call" --> Researcher[Researcher Agent<br/>Llama 3.3 70B]
        Nexus -- "Tool Call (MCP)" --> Calendar[Google Calendar / Meet]
    end

    subgraph "Knowledge & Tools"
        Researcher <--> VectorDB[(Weaviate Augmented DB)]
        Researcher <--> Search[Google Search API]
        DemoSpec <--> MockData[(Isolated Demo Contexts)]
        Calendar <--> GoogleAPI[(Google APIs)]
    end

    subgraph "Reliability & Privacy"
        Nexus -.-> RedTeam[Adversarial Red-Team]
        RedTeam -.-> Guardrails[Pydantic Validation]
        Guardrails -.-> Nexus
        API -.-> Fingerprint[SHA-256 Fingerprinting]
    end

Loading

---

Key Engineering Challenges & Trade-offs

1. Modularity vs. Latency

Multi-agent systems introduce coordination overhead. The shift to a leaner architecture and selective delegation reduces unnecessary inference steps.

---

2. Reliability vs. Flexibility

LLMs introduce non-determinism and failure modes.

A multi-layered approach was implemented:

• Retrieval grounding with confidence thresholds
• Adversarial validation (red-team logic checks)
• Structured outputs (Pydantic) to enforce correctness and ensure consistency

This significantly improved output reliability in evaluation environments.

---

3. Privacy & Session Management

Maintaining continuity without storing sensitive data:

• Salted SHA-256 hashing for identity tracking
• Summarized session memory ("conversation ghosting")
• No raw PII storage

---

4. Multi-Context Data Isolation

Ensuring separation between:

• Personal/professional knowledge
• Simulated/demo datasets

Achieved through strict collection-level isolation in Weaviate.

---

Results

• Reduced complex workflow execution time from hours to minutes
• Significant reduction in logical errors through validation, grounding and adversarial testing
• Improved cost efficiency via dynamic model routing and reduced redundant inference steps
• Reduced latency and token usage through architecture simplification

---

Tech Stack

Backend

• Python 3.11, FastAPI
• Google Agent Development Kit (ADK)
• Model Context Protocol (MCP)

AI / LLMs

• Gemini 3.1 Flash-lite
• Llama 3.3 70B (via Groq)
• Qwen 3:30B (via Ollama)

Data

• Weaviate (vector database)
• SQLite (metrics and anonymized tracking)

Frontend

• Next.js (React), TypeScript
• Tailwind CSS, Framer Motion

Infrastructure

• Docker, Docker Compose
• GitHub Actions
• Vercel

---

Notes

This is an actively evolving system focused on improving reliability, usability, and real-world applicability of multi-agent AI workflows.