Executive Summary

Prologue: Unearthing 2,322 Files

I wasn't trying to organize anything. I happened to be pulling a file list through the Dropbox API, and on a whim I typed .nb into the extension filter. Mathematica notebook files. When the results came back, the number stopped me cold.

2,709 in total. Subtract 236 from the [DS] team shared folder, 98 UST course materials, and 53 student assignments, and you get 2,322. Every one of them mine. From the first file in December 2011 to the last in March 2026. Fifteen years.

A Mathematica notebook is a peculiar format. Code, equations, graphs, and prose coexist in a single document. Like a laboratory notebook, you formulate the day's hypothesis, run the code, see the result with your own eyes, and jot down the next question. Each .nb file preserves that day's question like a fossil.

I counted these files by month of creation and drew a histogram. And there, laid out before me, was the story of my life.

Five distinct peaks emerge from the histogram. Summer 2013 (CLC research), 2018-2019 (the golden age of Code Painting), 2020-2021 (the founding era), 2023 (the LLM era), and the second half of 2025 (the LG project). The valleys between the peaks have stories of their own. And at the far right, where the graph bends vertically downward — that is where this essay begins.

The Researcher Years (2011-2017)

December 2011 — the first .nb file was born. In a lab at ETRI (Electronics and Telecommunications Research Institute). I was standing at the intersection of cameras and geometry.

CLC (Coupled Line Cameras) — a study on recovering camera parameters from a single image using nothing but the geometric properties of quadrilaterals. I solved the geometric constraints formed by vanishing points and diagonals of a rectangle through Mathematica's symbolic computation. The test image was the Dockland building in Hamburg. Its elegant parallelogram facade satisfied the geometric conditions perfectly. This research was presented at ICPR 2012 and 2014.

Mathematica was the essential instrument of this era. I expanded equations with symbolic computation, generated 3D visualizations on the fly, and recorded results directly in the notebook. I wasn't translating math from paper into code — the code itself was mathematics.

Around the same time, RSAR (Robotic Spatial AR) — Robotic Spatial Augmented Reality research — was underway, and in the Tangram Alive project, I solved tangram puzzle pose estimation in 0.2 seconds using Hausdorff distance. Thirty-four notebooks remain from camera calibration and Geometric Algebra research as well.

The first major peak appears in June 2013: 75 files. It was the climax of the CLC experiments. The daily routine of a researcher — splitting a single question open from hundreds of angles — reveals itself in the numbers. The 167 CLC experiment notebooks span from 2013 to 2024, a trace of one research theme expanding over more than a decade.

Painting with Code (2017-2019)

Two peaks leap out from the histogram before anything else. May 2018: 208 files. July 2019: 316 files. The annual maxima. What was happening during this period?

Code Painting. I had begun making pictures with Mathematica. The Shapes series (72 notebooks) composed geometric forms; Kinetic Art (27) set code in motion; Pixel Stack (18) built images by stacking pixels; Style Study (15) explored diverse visual languages; and Pebbles (8) shaped the forms of small stones. Some 140 notebooks were devoted purely to art.

It was the moment a researcher's tool became an artist's brush. Mathematica's functional programming and instant visualization created an environment where changing one line of code immediately changed the image. Every tweak of a parameter transformed the screen. The boundary between experiment and creation dissolved. Graphics[], Manipulate[], ColorFunction — these functions were my palette.

During the same period, Neural Style Transfer experiments were in full swing. A standout project cross-transferred the styles of Van Gogh's Irises and Gyeomjae Jeong Seon's View of Inwangsan Mountain After Rain. Western impressionism and Eastern ink-wash painting met inside a neural network. At the time it was artistic curiosity, but looking back, it was early work exploring the intersection of AI and art.

Alongside the art, experiments that would later become the seeds of a business were taking root in this same period.

Fly4ML was a delightful experiment. I generated synthetic fly images and used them to test Mathematica's Classify[] function. Training a classifier on self-made data — this was a small seed of the synthetic data business to come.

And then there was ModMan SLS (Synthetic Learning Set). Across 40 notebooks, I designed a pipeline in Mathematica for generating synthetic training data through Blender's physics-based rendering. This project, which began in 2015 and continued until 2025, was the seed of PebbloSim. I had no idea at the time that it would become the core technology of Pebblous a decade later.

Startup and Transition (2020-2023)

In November 2021, I founded Pebblous. On the histogram, this period reveals a fascinating pattern — the file count drops sharply, then rises again, a wave-like oscillation.

After founding the company, Mathematica's role changed at its root. It went from being an experimental tool for research papers to a prototyping instrument for building the early form of DataClinic. Visualizing class-level density distributions in image datasets, detecting outliers, designing quality metrics — all of this happened inside Mathematica notebooks. Every algorithm validation before productization is recorded in .nb files.

Just before founding the company, in August 2021, there was a 3-day Art Camp. Thirty-one notebooks poured out in three days. Ten a day. It was the final concentrated explosion of art and code. Looking back, it was the last deep immersion before crossing over into the life of an entrepreneur. The density of ten files a day says it all.

In 2022, a folder called [CLV] Research Management appeared. The moment Mathematica transitioned from a research tool to a business tool. I used Mathematica for managing research projects, data analysis, and business plan simulations. This folder would eventually hold 427 notebooks spanning 2022 to 2026, peaking at 235 in 2025 — the largest project folder in all of Pebblous.

Meanwhile, September 14, 2022 holds an unusual record: 119 files uploaded in a single day. It was a bulk upload of Code Painting works — pieces actually created between 2017 and 2021, now being organized into Dropbox. The 2022.09 spike on the histogram isn't new creation but a settling of the past, a kind of digital archiving.

In 2023, a month after the ChatGPT API launched, I created my first notebook that called an LLM from inside Mathematica. It was the year ML/DL-related notebooks peaked at 48. The marriage of Wolfram Language and LLMs opened new possibilities — yet it was also the moment when the limits of Mathematica alone began to come into view.

The Peak and the Turning Point (2025)

2025 was the absolute zenith of my Mathematica usage — and its final blaze. The LG Electronics PebbloScope project — geography, spheres, data visualization. From March through September, 189 notebooks were devoted to this project alone.

Over just three months, July to September 2025: 467 files. Twenty percent of all 2,322 files are concentrated in these ninety days. September 2025 alone saw 238 files — the second-highest monthly peak in fifteen years, after July 2019 (316). That translates to roughly eight notebooks a day.

And yet, at this very summit, I felt Mathematica's limits most acutely. Complex interactions, web deployment, large-scale data processing. Mathematica excels at exploration and prototyping, but a wall stands between it and production-grade applications. As the LG project scaled, that wall grew ever more visible.

The Plot Twist — When Words Became Code (2025.10-)

At the very moment the Mathematica histogram plunged off its cliff, something else was beginning. Vibe Coding. My first encounter with Claude Code.

Code that had been typed by hand, stitch by stitch, for fifteen years — gave way to code conjured by words. The instant the last bar of the Mathematica histogram vanished, the first post went up on blog.pebblous.ai. One era ended; another began.

When I pressed Enter instead of Shift+Enter, what changed wasn't a keyboard shortcut — it was the unit of thought. In Mathematica, you typed a function and executed it. In Claude Code, you speak an intention and review the result. The smallest unit of execution shifted from a function to an intent.

I haven't abandoned Mathematica. It remains irreplaceable for symbolic computation and mathematical exploration. But for everyday coding, web development, and data pipeline construction, Vibe Coding is overwhelmingly faster. The notebooks I used to create several times a day over fifteen years — their role is now fulfilled by conversations with Claude Code.

Epilogue — From Handcrafted Stitches to a Robot's Brushstrokes

The evolution of code can be told in three sentences.

The tool has changed. From Mathematica to Claude Code. But the question remains the same: "What can we see through data?"

The researcher who once reconstructed the geometry of quadrilaterals with CLC now diagnoses the quality of AI training data with DataClinic. The way of seeing the world through mathematical intuition hasn't changed. Read the structure hidden in data, visualize what that structure means, and from there, ask the next question.

These 2,322 .nb files are one person's intellectual trajectory. From geometry to art, from art to business, from business to AI. And threading through all of it, a single disposition — the attempt to understand the world through code.

That the histogram's last bar has fallen to zero is not an ending. It simply means the next question will be recorded somewhere other than an .nb file. Perhaps this very essay — fifteen years from now — will turn out to be the first bar of another histogram.

Why This Record Matters to Pebblous

There is a reason this apparently personal memoir appears on the Pebblous blog. The trajectory revealed by 2,322 .nb files explains where Pebblous came from.

The Origin of Synthetic Data: From ModMan SLS to PebbloSim

ModMan SLS — the experiment in creating synthetic training data through physics-based rendering, which began in a Mathematica notebook in 2015 — became PebbloSim a decade later. The trial and error captured in 40 notebooks evolved into the design philosophy of a Neuro-Symbolic Hybrid World Model. The conviction that data quality determines model performance first took shape in Fly4ML's synthetic fly image experiment.

The DNA of Data Quality Diagnosis: From CLC to DataClinic

The essence of CLC research was "recovering hidden structure (camera parameters) from incomplete data (a single image)." DataClinic does the same thing. It diagnoses quality problems in incomplete or contaminated data, identifies structural defects, and suggests remediation. The mathematical sensibility that once calibrated cameras through Geometric Algebra is woven directly into the design of data quality metrics.

Practical Questions Raised by a Tool Transition

The shift from Mathematica to Vibe Coding holds implications for Pebblous's clients as well. When an enterprise adopts AI, how should it transition its existing tools and workflows? How is the knowledge accumulated in fifteen years of .nb files transferred to a new environment? These are questions Pebblous confronts daily while building DataClinic.

Questions to Explore Next

The AI-generated .nb files that appeared in February 2026 raise new questions. How do we verify the quality of AI-generated code? When AI-created training data is used to train yet another AI, how must data quality standards evolve? These are precisely the challenges DataClinic must tackle in its next chapter.

Frequently Asked Questions

References