AI · SYSTEMS · STRATEGY
~/ai
AI in 2026 isn't a tool — it's infrastructure. The gap between organizations that ship AI systems and those that talk about shipping AI systems is a canyon, and it widens every quarter.
I work at the layer where models meet production: agent architectures that actually execute, inference pipelines that don't collapse under load, and cognitive automation that delivers compound returns instead of demo-ware.
The models are commoditizing. The [systems around them] are the moat.
Here's what most AI consultants won't tell you: I'm not advising from outside the building. I'm inside one. I run growth strategy for a multi-location restaurant group in DC while building the software that runs its daily operations. When I talk about AI in production, I mean the system my own team opens every morning.
And the honest version of what AI does to a business: it doesn't just cut costs — it consolidates work. The same team runs more locations. The owner gets hours back. The operation gets tighter without getting bigger. That's the return worth building for.
~/systems
Every system I design starts with the same question: what happens when this breaks at 3 AM? If the answer involves a human reading logs, the architecture isn't done.
Stack philosophy: choose the boring technology, then make it do something interesting. Postgres before vector databases. Bash before orchestration platforms. Monoliths until the split point is provable, not predicted.
The proof: CO-OPS, an operations platform I built solo and run in production for a two-location restaurant group. Opening verifications, closing checklists, role-gated workflows, multi-language from day one — used by real staff, every shift, since 2026. Not a demo. Not a pilot. The thing the restaurant runs on.
How it gets built matters as much as what gets built. Every project runs the same discipline: spec first, discussion before code, a living amendments doc where operational reality overrides the original plan, and a review gate before anything ships. When a line cook's workflow contradicts the spec, the spec loses.[Reality wins].
~/work
Custom operations platform for a two-location DC restaurant group. Checklists, opening/closing verification, role-based permissions across an 11-level hierarchy, full English/Spanish support. Built solo, Next.js + Supabase, in production daily use.
Recurring hours-reversion on a restaurant's Google profile, source unknown, costing walk-in revenue. Traced to a silent POS integration overwrite. Documented, validated, resolved at the source.
Third-party delivery commissions versus in-house economics. The analysis is done and the rollout is live — the numbers are still compiling. This one publishes with actuals, not projections.
~/strategy
Strategy in AI isn't about predicting which model wins. It's about building organizations that thrive regardless. The firms winning right now aren't the ones that picked GPT-5 over Claude — they're the ones whose architecture makes the model interchangeable.
I help teams move from AI FOMO to AI fluency. That means: auditing existing workflows for automation surface area, designing evaluation frameworks that measure what actually matters, and building the internal platform that turns AI capability into business velocity.
Example of what this looks like. A restaurant's Google Business Profile kept reverting its hours — costing walk-in traffic every time it happened. Everyone was guessing: Google bug, staff error, hack. I traced it to the POS provider's "Order with Google" integration silently overwriting the profile. Diagnosed the root cause, documented it, killed the failure mode. No report. No retainer. A problem that stopped existing.
That's the pattern for everything I take on: find the actual mechanism, fix it at the source, leave behind a [running system] — not a slide deck.
~/contact
If you're building AI systems that need to actually work — at scale, under load, in production — let's talk. If you're looking for a chatbot demo, there are plenty of agencies for that.
How this works: fixed-scope projects, per-project pricing, and you own everything I build — code, docs, the lot. First step is a short call where I ask about your operation and tell you honestly whether there's a system worth building. If there isn't, I'll say so.