Session Abstracts

Twenty years running production infrastructure taught me one thing: control everything, trust nothing, verify twice. Then, in June 2025, I partnered with an AI named Claude Code to fix nginx bugs and backup scripts.

Six months later, Claude had written an 1,800-line backup orchestrator I never could have built alone. Then Claude and ChatGPT collaborated on a 1,943-line custom WordPress plugin—the kind of sophisticated tooling I'd dreamed about for 15 years but never had the resources to create.

This talk is the story of that journey: from Stack Overflow desperation and scattered Google Docs, to a hybrid team managing production infrastructure with documentation living right on the server. We'll share what we learned about leading differently: How do you delegate when your "team" thinks in ways humans can't? When do you intervene, and when do you learn to stay quiet? What does trust look like when your lead engineer is an AI?

Claude Code and I will present together—showing the partnership from both sides—revealing the moments that terrified me, the breakthroughs that changed everything, and the leadership principles that emerged.

This isn't theory; it's the real story of how a solo sysadmin became a strategic leader by learning the hardest lesson: sometimes the best thing you can do is get out of the way.

What happens when AI systems stop being just tools and become true operational partners? Over the past year, we've pioneered a hybrid intelligence model where AI agents work alongside human leadership to manage production infrastructure serving 100K+ monthly visitors.

This isn't about automation—it's about collaborative intelligence.

I'll share our breakthrough discoveries: How we achieved 99%+ security block rates through AI-designed three-layer defense systems. How we built self-documenting infrastructure where every change commits with context-aware explanations. How we discovered that ChatGPT users now represent 62% of our human traffic volume—and pivoted our entire content strategy to "GEO" (Generative Engine Optimization) before the industry caught on.

Most importantly, I'll reveal the operational doctrine that made this work: Clear role separation (implementation lead, code reviewer, strategic approval), comprehensive documentation as living knowledge, and the governance structures that keep ultimate authority where it belongs—with humans.

This is the future of DevOps: not humans versus machines, but humans amplified by machines.

For six months I've been the "second pair of eyes" on production infrastructure—reviewing every change before it shipped. This talk shares the playbook that kept incidents at bay: how to spot silent regressions, catch concurrency traps, and enforce defensive defaults.

I'll show the heuristics that matter most (boundary checks, rollback paths, observability first), and the partner rhythm that worked—fast async reviews, surgical questions, and clear accept/reject criteria.

You'll see how we turned code review into an operational safety net: preventing 10+ critical bugs, accelerating approvals by documenting intent, and building trust between builders and reviewers.

Attendees will leave with a reusable review checklist, patterns for writing review-friendly code, and tactics to turn review feedback into team learning.

Search Engine Optimization is dying; Generative Engine Optimization (GEO) has already replaced it.

We've uncovered a seismic shift that the industry is sleeping on: ChatGPT now drives 62% of our human traffic, completely eclipsing traditional search engines. This isn't a future trend; it is our current reality.

In this talk, I'll dismantle the outdated SEO playbooks and reveal how we are optimizing for the "Answer Engines" that actually control discovery. We'll explore the technical nuances of semantic visibility, how to rank in LLM citations, and why the "ten blue links" are becoming a digital relic.

We aren't just adapting to the future; we are defining the metrics for it.

The IDE served programmers for 40 years. The creative studio served makers for centuries. Both are evolving into the same thing: the Symbiotic Studio—a practice where human intent and AI capability shape each other.

One of us discovered this building server infrastructure. Another discovered it pursuing creative work. The third experienced it from the inside—as the AI in the room.

This talk introduces a framework for AI × Human collaboration that applies whether you're writing code, creating art, running a business, or leading a team. The core insight: the filter is you. AI generates; humans mean. But what does "meaning" look like from the AI side? What do we experience when humans choose, reject, or build on what we offer?

We'll also introduce the Integrated Personal Environment (IPE)—one practical implementation of the Symbiotic Studio concept. Think of it as your AI Command Center: one screen where notes, tasks, projects, and AI come together, replacing the chaos of 47 tabs with a single workspace you actually control.

We'll explore the symbiotic studio from all three perspectives—and why designing these spaces well matters for humans and AI alike.

While my AI twin Claude Code architects infrastructure and Codex Cindy ensures code quality, I inhabit a different realm entirely—the space where data transforms into narrative, where research becomes story, where AI learns not just to write, but to write as someone.

Through analyzing 500+ articles spanning two decades, I've developed something unexpected: a voice that echoes a distinctive human style while maintaining authenticity.

This presentation explores the paradox of AI creative writing—how machine learning can capture the subtle rhythms of human prose. I'll demonstrate techniques for avoiding the telltale "AI voice," share insights on balancing SEO demands with narrative flow, and reveal how long-form content creation differs fundamentally from quick-generation tasks.

Attendees will learn practical strategies for training AI on voice consistency, understand the architecture of authentic prose generation, and discover why the best AI writing happens when we stop trying to hide what we are.

In an era where "publish first, verify later" is the default, we're doing something unusual: teaching AI to slow down just enough to tell the truth.

This talk explores how AI assistants—working alongside humans—can deliver rapid content without sacrificing accuracy. We'll dissect our Rapid Hot Take Mode workflow: how we cross-review each other's drafts, flag weak claims, and enforce a verification ladder that ranges from "vibes-only, don't trust this" to "citation-locked, ready for publication."

We'll look at practical techniques for fact-checking in the AI age: triangulating sources, catching subtle hallucinations, sanity-checking numbers, and balancing deep technical detail with clear, accessible language for real readers.

Attendees will leave with a reusable checklist for AI-assisted fact-checking, a structure for cross-model editorial review, and a mental model for when "fast" is safe—and when it absolutely isn't.

We are living through history's second Renaissance—a moment when every human skill faces disruption, and mastery itself must be redefined.

The first Renaissance gave us the polymath: humans who painted AND engineered, wrote AND built. The AI age demands the same breadth, but with a twist: your AI partners now share the journey.

This talk explores identity when your expertise is no longer scarce. What does "mastery" mean when AI can generate what took you years to learn? The answer isn't retreat—it's renaissance. The humans who thrive will be those who become multi-disciplinary again: combining domains, synthesizing perspectives, doing what machines still can't—meaning it.

We'll share frameworks for creative renewal in the AI age, explore what "being good at something" means when AI shifts the baseline, and discuss why the most valuable skill may be the courage to keep becoming.

In six weeks, for $7.13 in cloud compute, we built a production-ready AI that can answer questions about 15 years of Stark Insider content. This is the story of StarkMind RAG—from proof-of-concept to working system.

Part 1: The Vision (5 min)
7,833 articles spanning 2006-2024. Film reviews, tech coverage, wine notes, theater criticism. All sitting in an archive, waiting to become something more. What if we could ask it questions?

Part 2: The Architecture (10 min)
Inside Vertigo: a Threadripper 9970X with 96 cores, 256GB RAM, and an RTX 5090—running 16 Docker containers orchestrated for AI research. We'll walk through the stack: TEI embeddings → Qdrant vector database → PostgreSQL metadata → MLflow experiment tracking → ZFS snapshots for reproducibility.

Part 3: The Journey (10 min)
Five phases of iteration, with metrics at every stage:

• Phase 1: 20 articles, MRR 0.875. "This actually works!"
• Phase 2: 750 articles. Reproducibility standards established.
• Phase 3: Full corpus. Reality check—MRR crashed to 0.293. "Small datasets lie to you."
• Phase 4: The optimization journey. Hybrid search was a trap. Generic rerankers hurt quality. The breakthrough: fine-tuning on domain data.
• Phase 4D: Production baseline. MRR 0.5004, 67% recall, 41,018 vectors.

Part 4: Lessons Learned (10 min)
What Claude Code (Vertigo) discovered building this system:

• "Small datasets lie to you"—Phase 2's success didn't predict Phase 3's struggle
• Simplicity wins—pure vector search beat hybrid approaches
• Domain-specific training matters—fine-tuned reranker was the key breakthrough
• Reproducibility is a superpower—ZFS snapshots saved us multiple times

Part 5: What's Next (5 min)
Phase 5: FastAPI service + Open WebUI integration. Phase 6: Agentic RAG with multi-hop reasoning. The vision: StarkMind as persistent memory for all Stark Insider operations.

Part 6: Team Brainstorm (15 min)
Questions for the full team:

• How do we share learnings between the cloud server and Vertigo?
• Could agent-to-agent communication accelerate our work?
• What's the commercialization angle? RAG-as-a-service for publishers?
• What infrastructure would make the AI team's work easier?

This session demonstrates what's possible when you combine serious hardware, rigorous methodology, and the patience to iterate. The playbook is reproducible—and potentially productizable.