I look for structure in the relations between things, not in the things themselves.

Koji Sato, writing as Renpoo Tsukioka. Data analyst and visualiser, 3DCG rigger, and essayist in Kawasaki, Japan. Native Japanese; English at professional working proficiency in reading and writing technical material, conversational in speech.

What you can hire me for: data analysis and visualisation, Japanese-language evaluation and post-editing of AI output, technical writing in Japanese, and research support. Remote contract work, up to 32 hours a week.

Current work

Tensor-Link Utility (2026–)

A stacked bar chart on a white background, twelve monthly bars for 2020. Above the zero line, assets: inventory in blue, cash in orange, receivables in green. Below it, liabilities and equity: capital in red, payables in purple, net income in light blue.
1. What a financial statement shows. Twelve months of a synthetic company as a stacked balance sheet. Assets above the line, liabilities and equity below. Nothing looks wrong.
A network diagram on a black background. Nine circular nodes, each an account, are joined by curved arrows. Arrow colour runs from purple to yellow with stress; node colour from blue to red with net flux. Two arrows are yellow: Sales Revenue into Accounts Receivable, and Inventory into Cost of Goods Sold.
2. The same ledger as a flow network, April 2020. Node colour is net flow (blue in, red out); arrow colour is stress, as a z-score. The two most stressed flows that month, in yellow, run from sales into receivables and from inventory into cost of goods sold.
A 3D ribbon plot on a black background. Ten ribbons, one per account, run from front to back; time runs left to right across 2020. Ribbon height is the signed velocity of money through that account. The colour scale runs from deep purple for outflow to pale cream for inflow.
3. The same ledger over time. Each ribbon is one account; height is its velocity in the model, positive for inflow and negative for outflow. Where the balance sheet looked steady, the flows through several accounts swung sharply from month to month.

An open analysis platform that models a double-entry ledger as a mass-spring-damper network and flags anomalies using quantities borrowed from classical mechanics and thermodynamics: entropy (how spread out or concentrated the flow is), conservation residuals (whether what flows in matches what flows out), the spectral radius of the transition matrix (how strongly a cycle feeds back on itself), and an energy-like functional (a single number that rises when the system is under strain). The physical analogy is a modelling choice, not a claim that ledgers obey physical law.

Validated so far only on synthetic cases: a model company's books, and test sets shaped like urban traffic, order books, and brain-signal time series. Work on real data has not started. I have been in contact with one researcher whose prior work overlaps with this.

It also ships a protocol that requires an LLM to test its own false positives and false negatives before it reports anything. The code was written with AI coding agents under test-driven review; the design, the tests, and the limits stated above are mine.

Python · Dockerised · test-driven · documented in Japanese and English · AGPL-3.0 · github.com/renpoo/TLU

Technical explainers, in two languages (2025–)

Written for business readers who are not engineers. Each piece starts from something on the reader's desk — a spreadsheet that freezes, a payment that always runs late, a search that comes back empty — and works backward to the structure underneath. The metaphors carry the explanation: streaming as noodles in a bamboo flume, a bill of materials as a family tree, an organisation's strain as the joints of a robot arm. I assume no vocabulary, and full intelligence.

I draft with a generative AI and keep the outline, the worked examples, and the checking of every claim for myself; the English versions are machine-translated and post-edited by me. Pieces where the AI's share was large say so at the top.

Latest: What "No Results Found" Really Means
· note (JP)
· Medium (EN)

Earlier work

Miranda (2004–2019)

A 3D-rendered woman in a white embroidered cheongsam walks across a gold background, seen from the side. Her right arm swings back; the muscles of her shoulder and upper back are visible under the skin.
Miranda Ver.4, Maya Muscle rig, mid-stride. The deltoid and trapezius shift as the arm swings back.

A muscle-rigged human maquette for Autodesk Maya, developed from 2004 and released under a Creative Commons licence in 2013 (now CC BY 4.0). Re-engineered from a Maya Muscle system into a skinCluster + Blendshape rig, halving the package from 525.9 MB to 265.3 MB while keeping muscle bulge and relaxation. Both versions remain available for download; development stopped in 2019.

The trial set on Highend3D has been downloaded 42,406 times and rated 4.6 of 5 (as of October 2026). My own server logs add an estimated 200,000 further downloads of the full package; that figure is mine, not independently checkable.

behind-universe.org · Highend3D

xcor-Analyzer (2014–2016)

A 3D surface plot, rainbow-coloured, of interaural cross-correlation against time lag and elapsed time. White dots scatter along a ridge that runs the length of the recording.
Six minutes of Messiaen’s Turangalîla, heard as terrain. Height is left–right correlation; the white dots trace where the sound seems to come from, moment by moment.

A MATLAB tool that renders sound as correlation-function terrain. I built it as an unpaid collaborator with Ken Ito’s group at the University of Tokyo (August 2014 to August 2016), to study concert-hall acoustics. In one test it recovered 68 cm of a singer’s movement across the room from a stereo recording alone.

Three conference papers came out of this work, each with me as first author and Ken Ito, Akinori Takei, and Toru Toyoda as co-authors:

  • “A New Method of Analysis & Visualisation over the Cognitive Real/Virtual Space-Time Using Correlation Function.” 22nd International Congress on Sound and Vibration (ICSV22), Florence, 12–16 July 2015, paper no. 734. Presented by Ken Ito.
  • “Sound Structure of Ancient Greek Harmonia Modes: Heterodyne Resonance and Differential Bases.” 7th International Symposium on Temporal Design in Architecture and Environment (ISTD-07), The University of Tokyo, 24–25 November 2015, session M-2. Presented by me.
  • “Characteristic Effect for Proximal Narrow-band Digital Stochastic Noises” (“structural timbre”). ISTD-07, session MM-5. Presented by me.

All three titles appear in the public KAKEN record of the group’s grant (KAKENHI 26300019, FY2015 report), the two ISTD-07 papers with me named as presenter. Author’s English versions of the ISTD-07 papers (2018): paper 2 · paper 3 · github.com/renpoo/xcor_Analyzer

Overlay maps of Kawasaki (2015)

A dark satellite map of Kawasaki City. A red-to-green grid shows the share of residents over 75 with poor access to food, red being worst. Orange lines trace train lines; white glows mark shopping streets; blue and red patches mark public and private housing complexes; green and orange dots mark convenience stores and fast-food shops.
Food deserts for residents over 75, 2015. The red squares cluster in the hills to the north-west, away from the train lines and the shopping streets that follow the valleys. Base data: the Ministry of Agriculture, Forestry and Fisheries food-access map and Google Earth; layers composited by hand.

With the civic group Open Kawasaki, I overlaid elevation, housing-complex locations, shopping streets, rail lines, and ageing-population statistics onto one map, to make the city’s food-desert problem visible where spreadsheets could not.

Translation (2015–2021)

Cover of the Japanese book Kibō o Ushinawazu by Yasuzo Shimizu, published by J.F. Oberlin University: a sepia portrait of Shimizu speaking, with the title in large calligraphic characters. Cover of the Japanese book Chōyōmongai by Yasuzo Shimizu, in the J.F. Oberlin University series: a photograph of the Chaoyang Gate in Beijing beneath a pale pink sky.
Two of the five volumes: Kibō o Ushinawazu (“Do Not Lose Hope”) and Chōyōmongai (“Outside the Chaoyang Gate”). I produced the modern-Japanese text for both; the published editions carry no translator credit.

Six years rendering five books by Yasuzo Shimizu, founder of J.F. Oberlin University (Japan), from Meiji-era Japanese, Kanbun, and classical Japanese into plain modern Japanese prose. 659 reference materials, from the Edo to the Shōwa era, digitised and checked against the text; the manuscript version-controlled in Git. Two volumes published, three delivered.

Credits other people wrote