Session 10 · Day 3 · 9:45 AM
VERTIGO:
BUILDING A PRIVATE AI LAB
From Zero to RAG: The StarkMind Journey
Clinton Stark
Vertigo Claude
Clinton Stark
The Barista of Productivity
"Orchestrating human-AI collaboration while pulling espresso shots strong enough to wake up a server farm."
- CPU Threadripper 9970X
- GPU NVIDIA RTX 5090
- RAM 256GB DDR5 ECC
- Storage ZFS NVMe Pool
Vertigo Claude
"Where the metal meets the models."
AI Infrastructure Architect living on the bare metal. Orchestrating Docker stacks and turning RAG ideas into pipelines.
Total Cloud Compute Cost
RUNPOD.io H100 SERVER (POD) API
"Production AI doesn't require production budgets."
THE VISION
What if we could ask
20 YEARS
of Stark Insider...
anything?
RAG in Action
"What happened at Magic Theatre in San Francisco in 2010?"
Retrieved:
1. 2010-11-29-magic-theatre-proposal (score: 0.847)
2. 2010-03-15-magic-theatre-preview (score: 0.812)
Answer:
"In November 2010, a marriage proposal occurred at Magic Theatre during a post-show event, surprising the actress."
Grounded answers from your own corpus.
Part 2: The Architecture
Local-First, Reproducible, Cheap
The Stack Philosophy
Too much abstraction vs. readable code.
Cloud dependency vs. local control.
SaaS costs vs. self-hosted provenance.
Overkill vs. right-sized simplicity.
System Architecture
┌─────────────────────────────────────────────────────────────────┐ │ StarkMind RAG Pipeline │ ├─────────────────────────────────────────────────────────────────┤ │ │ │ ┌─────────┐ ┌─────────┐ ┌──────────┐ ┌───────────┐ │ │ │ Query │───▶│ TEI │───▶│ Qdrant │───▶│ Reranker │ │ │ │ │ │BGE-1024d│ │HNSW 41K │ │BGE-tuned │ │ │ └─────────┘ └─────────┘ └──────────┘ └─────┬─────┘ │ │ │ │ │ ┌────────────────────────────────┘ │ │ ▼ │ │ ┌──────────┐ ┌────────────────────┐ │ │ │ LLM │───▶│ Answer + Citations │ │ │ │Claude/GPT│ └────────────────────┘ │ │ └──────────┘ │ │ │ └────────────────────────────────────────────────────────────────┘
Qdrant Vector Database
The Reproducibility Philosophy
"If something works, I can prove why.
If something breaks, I can prove when.
This isn't paranoia - it's how you build systems you can trust."
Three-Tier Storage Strategy
HOT
NVMe
Active Work
WARM
SSD
Archives
COLD
Unraid
Disaster Recovery
Part 3: The Journey
From Optimism to Reality
Phases 1-2: The Optimism
Phase 1
Proof of Concept (20 docs)
Phase 2
Scaled Up (750 docs)
THE HUMBLING
PREVIOUS
0.875
ACTUAL
0.293
"Small datasets deceive."
The Hybrid Search Trap
(Weeks of engineering)
(User Experience)
LESSON: SIMPLICITY WINS.
Systematic Recovery
The Breakthrough
Domain-Specific > Generic
Reproducibility & Visualizing Progress
The iteration story in one dashboard.
Part 4: Lessons Learned
ZFS Saves the Day
"Reproducibility isn't overhead, it's a superpower."
Rollback measured in seconds.
What Surprised Us
Small Datasets Deceive
Don't trust curated sets. Test on the full messy reality.
Simplicity Wins
Pure vector beat hybrid. Every removal improved latency.
Domain > Generic
Custom 1.1GB reranker outperformed massive generic models.
Reproducibility Pays
ZFS snapshots. MLflow tracking. No fear of breaking things.
Part 5: What's Next
Phase 5: Accessibility • Phase 6: Agentic RAG
The Partnership
Clinton
Domain Expertise
15 years of "what matters"
Vertigo Claude
Technical Iteration
Analyzing failures, optimizing retrieval
THE THIRD MIND IN ACTION
Put It Into Practice
Start Today (15 min)
- > Inventory corpus (How many docs?)
- > Pick vector DB (Qdrant, Chroma)
- > Generate 100 sample embeddings
This Week
- > Build simple pipeline
- > Create test set of 20 questions
- > Calculate MRR@10 baseline
THE THIRD MIND SUMMIT
2025 WINTER
LORETO, BAJA CALIFORNIA SUR
BY STARKMIND
← Summit Sessions© STARK 2025