Maritime Domain Awareness · Internal demo
A good instrument doesn’t spend it. It stays dark until the signal is real — then shows you the one thing that matters, and lets you keep it.
The argument
Everything, equally loud. Thousands of contacts, every one shouting — watermarked, dense, exhausting. It spends your attention and hands back nothing.
Dark until it matters. Contacts that mean something, a sweep that reveals, a pin to carry the one you care about. You look once — at the thing that counts.
Why it matters
Let the agents run. The officer still needs one place to look — an instrument they open every shift and act from. That surface doesn’t disappear in the agent era; it just finally does its one job right: demand less, and spend your attention only on a real signal.
The method
You’ve heard it: AI hallucinates — makes things up. Fair, and true of raw output. Most people stop there and either ban it or trust it blindly. Both are wrong.
You don’t act on raw either. A rule fires when a shipment doesn’t add up — cheese out of the tropics, an entry that doesn’t match the bill. You built the checks that catch it. AI is the same craft: you don’t trust the output, you build the checks around it until what reaches you is verified, not guessed.
“We use AI” is what every contractor says. Here is what we mean when we say we know it — four things we actually do, on every job:
Every result runs against checks we wrote first. If a check can’t be made automatic and repeatable, the work doesn’t ship — the same bar as a rule that has to fire the same way every time.
We reproduce the failure first, then prove the fix kills it. “Looks right” is not a state we accept — not on a shipment, not from a model.
A second AI — a different engine — re-checks the first, blind to how it was made. Two independent sets of eyes, by construction, not courtesy.
Wording and data, it handles. Anything that changes a rule, a system, or a decision — a human owns it. On purpose. That line is the difference between a tool and a liability.
That rig — a clear goal, checks written first, an independent second opinion, a fallback when it fails, an audit trail of every step — is what we mean by a harness. Not a metaphor. It’s chain of custody around the AI, so what reaches the officer is evidence, not a guess.
And you’re looking at one. This page and the instrument behind it were built exactly this way — two people and an AI, under that discipline, no contractor bench. It runs the same every time, offline; the data that’s real is sourced, the data that’s simulated is labeled. Nothing is faked. That isn’t a claim about knowing AI — it’s the deliverable.
Why now — for years this discipline cost more than it saved: the tools needed too much hand-holding to trust. That changed. When the floor picked up Gemini 3.5-flash, the model got good enough that the harness finally pays for itself — the work runs, checks itself, and delivers. The method was always right; the tool caught up. The moment is real.
And who’s behind it
Between us, 26+ years inside the federal government — and 13 of them working the same problems, side by side, as the same two-person team. Not outsiders guessing at the mission. Not a rotating bench of contractors.
And it’s worth noticing who you’re buying from. For most vendors on the floor, AI that trims a contract’s headcount simply isn’t in their interest — their revenue is billable bodies, so a leaner team never helps their balance sheet. That’s no knock on them; it’s just the incentive they work under. We’re not on that model — we don’t sell heads. We’re the team plus a work buddy, aligned to the mission instead of to a roster.
The instrument
Self-contained · simulated contacts · built by two people and a work buddy.