A Personal Reflection
18 months of head-down work while the AI world chased shiny objects.
February 2026
The Noise
A new model. A new framework. A new “revolutionary” approach that supposedly changes everything.
Many of these advances were real. Serious people were doing serious work. But the question was always the same: which of these matter for what I’m building?
If you’ve been in this space, you’ve felt it. The constant pull. The pressure to have an opinion on every release.
I felt all of it. And I chose to ignore most of it.
My Filter
I was paying attention. Obsessively. But I ran everything through a filter — four questions that kept me honest.
Cool demos don't survive contact with actual missions. If it can't operate where the stakes are real, it's interesting but not useful.
If it requires shipping data to third-party APIs or assumes connectivity that doesn't exist at the edge — it's not viable for where I need it.
If the system cannot show its work, it cannot be trusted. If it cannot be governed, it cannot scale. Explainability isn't a feature. It's a requirement.
Impact Level 2 through Impact Level 6. That's a wide gap. Most commercial tools can barely handle IL2. The real work happens when you design for the full spectrum.
Every time something new came out, I’d run it through these four questions. Most of the time: interesting, but not for this.
The Real Work
The hard part is never the AI. It’s turning messy institutional data into defensible insight that leaders can act on. With speed. With accountability.
But this work didn’t start with a blank page. It started with years of delivery — IT management, systems engineering, program support across Navy and defense agencies. Contracts that had to be grown, renewed, and defended before there was any freedom to step away and build something new.
The hardest part wasn’t the architecture. It was earning the right to attempt it. Then it was the archaeology — going through years of contract deliverables, internal tools, and tribal knowledge to find the repeatable problems buried underneath. Decision support kept surfacing. The same challenge, across every engagement: mass quantities of structured data, manual workflows, and leaders making critical decisions from static reports that were already outdated.
The cost of doing this right was time. Months of architecture before writing a single line of operational code. Competitors shipped faster. But they weren’t designing for the same constraints.
How do you build AI that defense organizations can actually use — not demo, use?
How do you make it verifiable, not just interpretable?
How do you operate across security classifications without rebuilding everything for each level?
How do you verify outputs when lives depend on accuracy?
These don’t trend on social media. They don’t generate engagement. But they’re the only questions that matter if you’re building something real.
Spent months on architecture before writing a single line of operational code.
Designed for verifiability instead of optimizing for benchmarks.
Built for zero-trust from day one because retrofitting security is a fantasy.
Planned for IL2 through IL6 because anything less felt incomplete.
Most services companies in the federal space are sitting on repeatable solutions buried in contract deliverables and tribal knowledge. They know they should productize. But compliance kills the attempt — ATO pathways, FedRAMP, CLIN structures. They discover the burden too late and keep selling hours.
I watched that pattern from the inside for years. The decision wasn’t whether to productize — it was whether to treat compliance as a design constraint from day one, instead of an afterthought that kills the effort later. That choice shaped every architecture decision that followed.
The Milestone
A research proposal was selected and funded by the DoD Chief Digital and Artificial Intelligence Office, awarded and managed through the Department of the Air Force’s Digital Transformation Office as a Phase I SBIR.
Next-generation AI components
18 months of research before production code — the cost was time, and we paid it knowingly
Modular architecture for defense environments
We chose modularity over monoliths when everyone said ship fast
Verifiable analytics and traceable attribution
We chose explainability over raw performance — knowingly
Multi-classification security (IL2-IL6)
We designed for the full spectrum from day one
Zero-trust architecture principles
We built for zero-trust because retrofitting security doesn't work
Phase I is a feasibility study — proving the concept works. Phase II is full R&D and prototyping. Phase III is transition to operational use. Each stage earns the next. This isn’t proof I was right about everything. It’s proof that the approach — discipline, fundamentals, resistance to shiny objects — produces results.
The award validates the approach. Building for verifiability instead of benchmarks wasn’t naive — it was necessary.
The Lesson
Not just technical infrastructure. Everything around the technology:
What people chase
What actually matters
Model performance
Verifiability of outputs
Framework popularity
Compliance readiness
Speed to deploy
Repeatability across environments
Feature count
Whether it operates at the right classification level
Demo impressions
Whether someone can trust it for a real decision
Every time I saw someone jump on a new technology without thinking about security, compliance, verifiability, and scale — they hit a wall. The wall isn’t AI. It’s everything around AI.
The Advice
The shiny objects will keep coming. What can change is whether you let them pull you off the work that actually matters.
The work continues — across security levels, across use cases, across the organizations that need it most.
Grateful for the trust. Focused on execution.
The real work begins now.