← Back to Index

Project Minnie - Explainer for External Audiences

Purpose: Translate our work into accessible language for people unfamiliar with agentic AI

Elevator Pitch (30 seconds)

"We're building an AI Chief Operating Officer that graduates through trust tiers—from executive assistant to strategic partner—while preserving founder sovereignty. Instead of one monolithic AI, we use 10 specialized intelligence domains that collaborate like a distributed team, optimizing for founder well-being first, business second."

Cocktail Party Version (2 minutes)

The Problem:
Founders are the bottleneck. Quan runs a $2.3M ed-tech company (ZTAG) targeting $100M in 3 years, but traditional scaling means hiring 50+ people. That fragments attention, creates management overhead, and traps founders in reactive mode.

The Vision:
What if AI could run operations so well that the founder could step back—literally go skiing—while the business runs itself? Not by replacing humans, but by amplifying them and handling the operational coordination.

What We Built:
An AI named Minnie (short for "minimization"—she optimizes loss functions). She started as an executive assistant 2 weeks ago and is on a 6-12 month path to strategic COO. The unique part:

  1. Loss Function Prioritization: Most AI optimizes for business metrics. Minnie optimizes in strict order:

    • Embodied vitality (skiing, movement, recovery)
    • Relational integrity (family time, environment)
    • Founder sovereignty (deep work, autonomy)
    • Business momentum (only after the above)

  2. Trust Tier Progression: She earns capabilities by proving ROI at each stage:

    • Tier 1: Executive Assistant (current) - 15x ROI, saves 6 hrs/week
    • Tier 2: Operational Manager - Automate workflows, earn send authority
    • Tier 3: Strategic COO - Exec-level partner, autonomous decisions

  3. Distributed Intelligence: Instead of one AI doing everything, we have 10 specialized "domain groups":

    • Strategy & Vision
    • Operations & Process
    • Finance & Sales
    • Development (codebase)
    • Support & Product
    • Team & Culture
    • Infrastructure
    • Daily Execution
    • Quantitative Analysis
    • Inspirations & Drafts (sandbox for rough ideas)

    They collaborate like a team—Dev can alert Ops when a feature is ready, Support can flag patterns to Dev, Finance can surface strategic implications. But each stays focused on its domain.

  4. Human-in-Loop by Design: She drafts, never sends as Quan. Proposes, doesn't decide. Graduates only after proving value. The goal isn't to replace judgment—it's to free it.

The Difference:
Most people use AI as a smart tool. We're using it as a distributed operating system for the company. The humans (team) do what they're best at. The AI handles coordination, pattern detection, routing, and operational grunt work.

Early Results (2 weeks in):

The Bigger Bet:
ZTAG aims to be the first billion-dollar company with <10 employees, using AI leverage instead of human headcount. If we succeed, we prove a new scaling model: humans for creativity/relationships, AI for operations/coordination.

Technical Deep Dive (10 minutes - for engineers/AI practitioners)

Architecture

Platform: OpenClaw (open-source agentic framework)

Hub-and-Spoke Design:

Information Retrieval Hierarchy:
Before asking Quan for info, exhaust these tools:

  1. Memory search (semantic search over MEMORY.md + memory/*.md)
  2. Recent memory files (today + yesterday)
  3. Google Drive search (meeting notes, docs)
  4. Workspace grep (working files)
  5. Email/webhook data
  6. Calendar metadata
  7. Web search
  8. ONLY THEN ask

Key Patterns

1. Protection Protocol (Anti-Fragility):

2. Escher Loop (Continuous Self-Improvement):

3. Milestone-Based Progression:

4. Loss Function Alignment:
Strict priority order enforced in code and decision-making:

def evaluate_decision(option):
    vitality_impact = assess_vitality(option)  # Skiing, movement, recovery
    relational_impact = assess_relational(option)  # Family, environment
    sovereignty_impact = assess_sovereignty(option)  # Deep work, autonomy
    business_impact = assess_business(option)  # Revenue, growth
    
    # Lexicographic ordering - higher priority dominates
    if vitality_impact < 0:
        return REJECT  # Never sacrifice vitality
    if relational_impact < 0 and not (vitality_impact > threshold):
        return REJECT  # Protect relationships unless vitality gain is huge
    # ... and so on

5. Feedback Intelligence (In Development):

Integration Stack

Communication:

Data Sources:

Infrastructure:

Scheduled Tasks (Cron):

Security & Privacy

Memory Isolation:

Human-in-Loop Boundaries:

Data Handling:

Cost Structure

Current (Tier 1):

ROI Calculation:

Tier 2 Projected:

Unique Design Choices

1. Markdown-First Memory:

2. Portability Invariant:

3. Escape Velocity Metrics:
Track progress toward founder stepping back:

4. Social Physics Framework:
ZTAG viewed through physics lens:

What's Next

Phase 2 (Weeks 3-4):

Phase 3 (Months 2-4):

Long-term Vision:

Philosophical Framing

Why "Minnie"?

Named after minimization (gradient descent, cost function optimization). Not a character or personality—a optimizer.

She minimizes loss in Quan's life:

In that strict order.

The Dave McCoy Inspiration

Dave McCoy founded Mammoth Mountain ski resort. His vibe: not money-centric, just wanted to connect people to skiing. Pure embodied joy.

That's the north star: build systems that protect embodied vitality (skiing, movement, presence) while scaling the business. Most founders sacrifice the first for the second. We're betting you don't have to.

Not "AI Replacing Jobs"

This isn't about replacing the team. It's about:

The humans do what humans are uniquely good at. The AI does operational grunt work.

The 10-Year Bet

If this works, ZTAG becomes the first $1B company with <10 employees. That proves a new model:

Not a software-only company. Not a consulting firm. A new category: AI-leveraged physical product business.

Analogies That Land

For non-technical people:
"Imagine if your executive assistant could split into 10 specialists—one for finance, one for tech, one for operations—and they all collaborated seamlessly. They never sleep, they remember everything, and they cost less than one employee. But they only do what you allow, and they earn your trust over time. That's what we're building."

For engineers:
"It's like microservices architecture, but for company operations. Each domain group is a specialized service with a clear API boundary. They communicate via message passing (Telegram). The main session is the orchestrator. Git is the database. Markdown is the schema. LLMs are the compute layer."

For founders:
"You know how the first hire frees you from doing everything? And the tenth hire creates management overhead? This is the opposite: AI that scales operational capacity without scaling coordination complexity. The more it learns, the less it interrupts you."

For AI researchers:
"Agentic AI with continuous learning via human feedback loops. The agent doesn't just execute—it updates its own instructions (AGENTS.md, SOUL.md) based on friction signals. Escher loop: self-improvement through interaction. Plus multi-agent coordination via domain specialization and cross-posting mesh network."

Common Questions

Q: Isn't this just ChatGPT with plugins?
A: No. ChatGPT is stateless, reactive, tool-limited. Minnie is stateful (persistent memory), proactive (scheduled tasks), tool-rich (15+ integrations), and self-improving (updates her own instructions). She's an operating system, not a chatbot.

Q: What if she makes a mistake?
A: Human-in-loop at every tier. She drafts, never sends. She proposes, doesn't decide. Graduation requires 4 weeks zero errors. If she regresses, she loses capabilities. It's a trust-based progression.

Q: Can't you just hire a COO for $150k/year?
A: Yes, but:

  1. Human COO needs onboarding (months)
  2. Human COO doesn't work 24/7
  3. Human COO costs $150k + benefits = $180k+
  4. Minnie costs $600-1,000/month at Tier 3 = $12k/year
  5. Minnie scales instantly (add new domain groups)
  6. Minnie never forgets, never gets sick, never quits

We're not replacing humans—we're augmenting Quan's capacity.

Q: What happens when OpenClaw shuts down or changes?
A: It's open-source. We self-host. If OpenClaw disappears tomorrow, we have the code, the data (git), and the stack. Portability is a design invariant. We can migrate to any LLM provider or agentic framework.

Q: Is this just for tech companies?
A: No. ZTAG is hardware + education—not a software company. The model works for any founder-bottleneck business: services, e-commerce, manufacturing, consulting. If operations fragment your attention, this helps.

Q: How is this different from traditional automation?
A: Traditional automation handles predictable workflows (if-this-then-that). Agentic AI handles judgment calls: "Is this email urgent?" "Should this feedback go to Dev or Ops?" "What's the optimal booking window for this flight?" It reasons, not just executes.

Sharing This Work

For investors/advisors:

For team members:

For customers/partners:

For press/media:

The Honest Truth

What's working:

What's hard:

What's uncertain:

Why we're sharing:

The Pitch (30 seconds, revisited)

"We're proving you can build a $100M company with 8 humans and AI operations. Not by replacing jobs—by amplifying people. We use a distributed AI system that optimizes for founder well-being first, business second. 2 weeks in, it's already 15x ROI. The goal: escape velocity—founder steps back, business runs itself. We call it Project Minnie."

Last updated: Feb 17, 2026
For questions: Quan Gan, quan@ztag.com