On thirty years of
convergent preparation

People sometimes ask how the SciPhAI platform came to exist — a new class of LiDAR architecture, a decade-spanning mathematical framework for machine intelligence, two patent families filed years apart but pointing at the same thing. The honest answer is that it did not arrive suddenly. It was prepared for over thirty years, from three directions at once, until the directions converged and the answer became, in retrospect, obvious.

I want to explain that preparation — not as biography, but as argument. Because the preparation is the reason the platform is what it is, and understanding the preparation is the clearest way to understand why the technology works the way it does and why no one else has built it.

I completed a Master of Engineering in a year and a Ph.D. in under eighteen months at the University of New South Wales in Sydney — eighteen papers, mostly as first author, across various linear and nonlinear photonic device physics, nonlinear optics, optical logic gates, optical switching, optical amplifiers, pulse compressors, electromagnetic theory, and information theory. The breadth was deliberate. I was building a foundation, not specialising in one narrow channel.

Those years produced a number of substantial contributions. One I am still proud of: I proposed, analyzed, and experimentally demonstrated multilevel soliton communication systems — a novel communication system and method for ultra-high-speed optical links with high figure of merit — published in the IEEE Journal of Lightwave Technology in January 1997.

At the same time, I never stopped circling a different question entirely: how do we come to know anything, and why is the universe so faithfully described by mathematics? A theoretical model predicts something; and when its founding assumptions are sound, nature attests to it. Those two preoccupations — how information moves through the physical world, and how a system comes to know — would stay with me for the next thirty years.

Underneath both ran a third thread, easy to miss: information theory. Telecommunications at the level I worked — ultra-high-speed fiber backbones, OC-48 access, dense wavelength multiplexing, the information limits of physical channels — is applied information theory: Shannon's entropy, channel capacity, mutual information. Spend two decades working at the theoretical limit of what a physical channel can carry and it trains a particular way of thinking about what information actually is — a habit of mind that would shape the knowledge framework far more than any single device ever did.

In 1996 I founded Peleton Photonic Systems Inc. The idea was to build metropolitan-area optical networks delivering OC-48 — 2.5 gigabits per second — to residential subscribers. This was the year-2000 era. The ambition was real: multi-wavelength laser sources, external modulators, wavelength-division multiplexed transmission, all-optical routing. We raised US$22 million and built the technology.

That first company also produced my first patent, filed in 1996: a method for multiplying the number of communication channels a limited set of sources can carry, by giving each channel its own distinct “photonic pattern.” A small thing to point to now — but it was the first time I reached for a structured, per-element representation to solve a problem, the same instinct that, a decade later, would become the core of the knowledge framework.

What Peleton taught me was not primarily optical physics — I already had that. It taught me the discipline of making photonic technology perform reliably in the field, end to end, under real-world conditions. The gap between a laboratory demonstration and a deployed system is where most optical startups fail. Closing it requires a different kind of thinking than the physics does — systems thinking, manufacturing thinking, the discipline of understanding failure modes before they occur.

In 1998, while Peleton was still operating, I co-founded Zenastra Photonics Inc. — focused on the fabrication of silica-based passive and hybrid photonic integrated circuits. Zenastra raised US$44 million. The work lived at the fabrication layer: planar lightwave circuits and the hybrid integration of active components onto passive substrates.

Peleton operated until 2007; Zenastra until early 2002, caught in the 2001–2002 telecom capital collapse — not the dot-com crash. The technology was never the problem. Between the two, I had built — as founder, CEO, and technical lead — one systems-level photonic company and one device-fabrication company, in the same decade that I was developing the mathematical framework that would become the intelligence layer of the platform.

The question that occupied me alongside the photonics companies — and increasingly after them — was a different kind of question. Not how to transmit light. How does a system come to know something?

Beginning with a provisional patent in 2008, I built a mathematical framework for how a machine can represent knowledge: as a structured, ordered account of how the parts of any body of information relate to one another across levels of detail — and, in the later work, of which of those relationships are causal rather than merely correlated.

At its core was a measure I derived from information theory: the conditional occurrence probability (COP) of one component of a body of knowledge or data given another — priority 2009 — which makes the information content of any composition analytically computable. Where Shannon asked how much information a channel can carry, this asks how much a composition carries about its participants: the same lineage, a different question. And the rows and columns of the matrices it produces — association strength and conditional occurrence — are, in effect, vector representations of concepts: word and concept embeddings, years before that name existed.

Over the years from 2009 onward this became more than twenty issued US patents, applying the same core framework to knowledge discovery, semantic processing, document and genomic analysis, and artificial intelligence. The specific methods evolved across the series; the underlying idea — that understanding is structured and causal, not a pile of statistics — stayed constant.

By 2011 the framework was running as a live public service — a conversational knowledge-discovery system anyone in the world could query, assembling a body of knowledge in real time and answering across several modes: novel, informative, consensus, causal. We called it Mr. SCI. The term "retrieval-augmented generation" was coined about nine years later.

Two further applications — one extending the framework to physical sensing and causal reasoning, one on the sensing hardware itself — are pending, and I come to them next.

By 2019, two decades of photonics and a decade of framework development had arrived at the same question from opposite directions.

From the mathematics side: the framework could, in principle, work on any body of data — text, genomic sequences, financial records, sensor streams. But what kind of sensory data would give a machine the richest possible picture of the physical world? A conventional LiDAR point cloud — a set of independent range measurements at a set of angles — is, informationally, a sparse and causally impoverished thing. It tells you distances. It does not tell you how those distances relate, which objects produced which returns, or how the scene is structured. A model built on such data can only ever be shallow.

From the photonics side: I knew, from twenty years of building optical systems, that the conventional LiDAR architecture — many independent laser-and-detector signal chains — was expensive, complex, fragile, and essentially unchanged from its origins. And I knew, from those years of nonlinear-optics work, how much structure a single propagating light pulse can carry when the medium is designed for it. An intuition that had lain dormant for twenty-five years suddenly had a purpose.

The result is a new sensing architecture in which the optics are designed so the scene reports itself in a structured, causally traceable way — every return tied to what produced it — with no moving parts and no external beam steering. It runs in two complementary modes: one tuned for wide, all-weather coverage, the other adding velocity and longer range. Both come from the same physical design.

The intelligence framework was filed in July 2019. It described a complete path from physical sensing to causal, interpretable machine decision-making, and named LiDAR, cameras, and radar as its sensory sources. Five years later, in October 2024, the sensing hardware was filed — the architecture built to produce exactly the kind of grounded, structured data that framework needs. The hardware was not conceived after the theory. It was motivated by it.

There is one more thread worth naming, because it comes from the same cross-domain habit that produced everything else. Much of my early work was about how energy and information transfer between two systems brought into contact — and how efficient that transfer is depends on how well their internal structures match.

The same intuition, it turns out, describes how knowledge transfers between a human and a machine. Efficient transfer requires that the machine's internal representation be compatible with the structure of human understanding. A machine that represents knowledge in a structured, causal way can meet a human expert on common ground. A machine that represents knowledge only as statistical weight over words struggles to, however fluent it sounds.

That was one of the things that convinced me the framework was genuinely foundational — not a useful engineering trick, but a description of something real about the relationship between physical systems and knowing systems.

I want to be precise about what I am and am not claiming. The idea that intelligence is grounded in physical reality is not mine alone. Wiener alluded to it in 1948. Brooks circled it in 1986. Lakoff and Johnson named the phenomenon in language. Gibson described direct perception. They asked the right question.

What I am claiming is this: I gave an engineering answer. A specific, patented, physically realizable answer to the question those traditions raised but did not resolve. What must a sensor produce for a machine to genuinely understand its environment? What mathematical structure must the machine's knowledge take for its decisions to be rational, interpretable, and causally grounded? How does knowledge transfer efficiently between a human expert and a machine?

The answers are in the patent record, dated. The preparation that made those answers possible took thirty years. The vocabulary arrived in 2025. The work did not wait for it.