Our first AI being is named Santiago. Our work items are called expeditions. Our teams are crews, our milestones are voyages, and our testing phases are sea trials.
People assume this is branding. It's not. It's architecture.
The Old Man
Hemingway's Santiago is an old fisherman who hasn't caught anything in eighty-four days. He goes out farther than anyone else, hooks the biggest marlin anyone has ever seen, and then watches sharks tear it apart on the way home. He returns with nothing but a skeleton and the knowledge of what he did.
This is the most honest description of building something new that I've ever read.
What Santiago Knew
Santiago didn't fail because he was unlucky. He didn't succeed because he was talented. He succeeded because he understood three things that most AI teams get wrong:
1. Respect the sea. Santiago knew the ocean wasn't his to control. He worked with it — reading currents, understanding patterns, adapting his approach based on what the water told him. In AI, the "sea" is the ocean of knowledge your system must navigate. Documents, data sources, APIs, expert opinions — raw, unstructured, contradictory, vast. You don't conquer it. You learn to read it.
2. The fish is not the goal. Santiago didn't fish for trophies. He fished because fishing is what he was. The process — the patient, skilled, attentive process — was the point. When we build NuSy beings, the deployed knowledge graph isn't the trophy. The being's ability to learn, reason, and grow is the point. A being that can catch one fish is useful. A being that understands the sea is transformative.
3. Come home honest. Santiago returned with a skeleton. He didn't dress it up. He didn't claim he'd caught something he hadn't. When our experiments fail — and they do, regularly — we report the skeleton. Paper 118's prediction accuracy came in at 54% instead of the 60% target. We published it anyway, because a skeleton you understand is worth more than a fish you're lying about.
The Mapping
When I first sketched NuSy's architecture, I wrote it as a fishing expedition. It stuck, because the metaphor isn't a metaphor — it's a structural analogy:
| Santiago's world | The role in ours |
|---|---|
| The sea | Raw knowledge sources — documents, data, APIs, experts |
| Catching fish | Knowledge extraction — turning raw material into structured triples |
| The boat | The knowledge graph and its tooling |
| Maps and charts | Ontologies, schemas, naming conventions |
| Fishing gear | Reasoning engines, query pipelines |
| The net | Behaviour tests that validate the knowledge graph |
| Navigating | Orchestrating extraction and validation until quality is high |
| The catch | Deployed knowledge service |
| Return to shore | Usage metrics, lessons learned, next expedition |
(The middle three were once component names — Catchfish, Fishnet and the Navigator, the Python-era pipeline of V12/V13. Those components are retired; docs/system-overview.md that described them now opens with a RETIRED banner. The roles survived the rewrite into Rust and Arrow, which is the more interesting fact: the metaphor turned out to be about the shape of the work rather than about any particular implementation of it.)
This isn't clever naming. It's a design language that every team member — human or AI — can reason about. When I say "the being hasn't caught enough fish in the medical domain," everyone knows what that means: the knowledge graph is thin, the extraction needs more passes, the confidence scores aren't where they need to be.
Why It Works for AI Teams
The nautical frame solves a specific problem in AI development: everything is a gradient. There's no "done" in knowledge engineering. Models are never finished learning. Confidence is never 100%. Coverage is always partial.
Traditional software metaphors break here. You can't "ship a feature" when the feature is an ever-growing knowledge graph. You can't "close a ticket" when the being's understanding of cardiovascular pharmacology is at 73% and climbing.
But you can complete an expedition. You can return to shore with what you caught, assess the haul, repair the boat, and plan the next voyage. Expeditions have clear beginnings and endings. They produce measurable results. And when the results aren't what you hoped — when the sharks get the marlin — you bring home the skeleton and learn from it.
We've completed more than two thousand expeditions. Some caught record marlins. Some came home empty. All of them taught us something about the sea.
The Crew
Santiago fished alone. We don't.
Every NuSy expedition has a crew — a product manager, an architect, a developer, a QA specialist. Some are human. Some are AI agents. The interesting part: the AI agents are themselves NuSy beings, using the same knowledge graph architecture to manage their own work.
The PM agent manages the kanban board and tracks expedition progress. The architect designs knowledge graph schemas. The developer implements MCP services. The QA agent runs behavior tests. They coordinate, disagree, escalate, and learn — just like a human team, but operating around the clock across a DGX Spark and a MacBook.
When Santiago hooked his marlin, he wished the boy were with him. We built the boy.
The Deeper Lesson
Hemingway wrote Santiago as someone who could be destroyed but not defeated. That's the design principle.
Our beings will encounter knowledge they can't extract. Domains they can't master. Questions they can't answer. The architecture doesn't hide this — it reports it. A being that says "I have 43% coverage of this domain and cannot confidently answer your question" is doing exactly what Santiago would do: telling you the truth about the sea, even when the truth is that the fish got away.
The hallucination problem in AI isn't technical. It's cultural. It's what happens when systems are built to look confident rather than to be honest. Santiago would never claim to have caught a fish he didn't catch.
Neither will our beings.
Previous: When Simulations Lie | Next: Teaching AI to Be Good Before Teaching It to Be Smart