Skip to content

Personal site — vol. 01

Curious by default.
Building to understand.

SOFTWARE · SYSTEMS · RESEARCH · EXPERIMENTS

I like taking systems apart to see what they're really doing — market data, desktop environments, local models, the tools in between. This site is the working notebook: research still open, prototypes in flight, and the questions that outlasted their answers.

systemslinuxsoftwareresearchexperimentscuriosityFIG. 00 — ORBITS OF ATTENTIONN = 5SCALE — ARBITRARYREV. 001

fig. 00 — what the work tends to orbit

Selected work
03
Experiments
ongoing
Answers
pending
01 / Selected work

Selected work

Three entries big enough to have history. Each one states the question it's actually chasing; statuses are honest, and where something is unsolved, it says so out loud.

01

Market Research Lab

Research · Ongoing

An ongoing research program asking one narrow question: when a market-data model is given more information — a forecasting signal, filings, fundamentals, news — does anything genuinely add out-of-sample signal?

The question

Most “more data improves the model” stories are backtest stories. The question is whether each information source earns its place under an evaluation that respects time — and if so, which ones, and by how much.

The approach

Built on Microsoft's Qlib, with LightGBM establishing the market-data baseline. Four configurations are compared under one time-aware research framework: point-in-time alignment, walk-forward evaluation, and out-of-sample measurement of each source's contribution. The exact recipes — blends, features, lags — stay private; the design is the public part.

One detail

The modes are compared, never stacked: a more complex configuration is not assumed to be a better one. Every feature carries a “knowable by” timestamp so no evaluation window can read tomorrow's filing today — and each mode faces the same walk-forward discipline as the baseline before any claim is made.

Stack

Python · Qlib · LightGBM · Kronos-based forecasting signal · SEC-related filing data · company fundamentals · news-derived context

※ Ongoing research; historical experiments only. No returns, rankings, or winning mode are claimed until verified out-of-sample evaluation supports them. Nothing here is investment advice or a trading system.

INFORMATION-SOURCE MATRIX — DESIGN, NOT RESULTSmarket datakronos signalfilings+ fundamentalsnews contextABCDALL FOUR MODES → THE SAME WALK-FORWARD, OUT-OF-SAMPLE EVALUATIONCOMPARED, NOT STACKED
fig. 01 — what each mode is allowed to see. Complexity is not assumed to help; every configuration faces the same time-aware evaluation.

The experimental matrix

compared, not stacked

Four configurations. Different information sources. One question: what genuinely adds out-of-sample signal?

AMarket Baseline

Reference configuration

Establishes the reference point: market-derived signals with LightGBM in the Qlib pipeline — the line every other mode is measured against.

asks — How much predictive signal does the market-data baseline provide before anything else is added?

Reference configuration

BForecast-Augmented

Forecasting-signal experiment

Pairs the baseline with a signal generated by a pretrained time-series forecaster, Kronos — conceptually a second opinion on where the series is heading. The blending recipe stays private.

asks — Does an independently generated forecasting signal complement the baseline, or add nothing it doesn't already know?

Forecasting-signal experiment

CCompany Context

Running · evaluation pending

Enriches the market view with SEC-related filings and company fundamentals — regulatory and financial context, deliberately without news features. The feature inventory and alignment rules stay internal.

asks — Does company-level context carry predictive information beyond market-derived signals?

Filings & fundamentals experiment

DNews-Enriched Context

Planned · awaiting evaluation

Builds on the same research idea as Mode C, adding time-aware news-derived context on top of market, filing, and fundamentals information.

asks — Does news-derived context contribute incremental information once company context is already present?

News-enriched experiment

02

Slate

Prototype · In progress

A local-first desktop assistant built for Linux — designed to feel like part of the operating system, not another chat window begging for attention.

The question

What would an assistant feel like on Linux if it were designed as a system utility — quiet by default, keyboard-first, aware of its context — instead of a messaging app that happens to live on a desktop?

The approach

Linux-only by conviction: the desktop-integration questions Slate asks only have honest answers on a system you control end to end. Tauri, Vue and Pinia with SQLite for history; models served through Ollama so everything stays on the machine. An earlier iteration explored QML and CMake.

One detail

The unsolved part is context: an assistant that knows which workspace you're in is more useful — and a more delicate idea. Getting the line between “aware” and “surveillance” right is the actual project; it is open.

Stack

Tauri · Vue · Pinia · SQLite · Ollama · QML/CMake (earlier iteration)

no public repository yet — ask me about it

slate — local sessionlinux · 0 network calls

⌘k — ask anything

ask anything — it stays on this machine

esc — clear · ⌘h — history · voice input — planned, accessible by default

model — local · ollamalinux only · built for this desktop
fig. 02 — interface direction as a concept sketch, Linux only. The real thing replaces it when it has earned a figure of its own.
03

Ryoku

Ongoing · Personal infrastructure

A personal Linux workspace assembled piece by piece around Arch and Hyprland — fast, quiet, and configured to exactly one person's taste.

The question

How much of the way a computer feels is the system underneath — and how much is the hundred small decisions stacked on top of it? The only honest way to find out is to make every decision myself.

The approach

Arch Linux with Hyprland for tiling and workspaces, Waybar for a status line that reports instead of performing, and dotfiles under version control so the environment can be rebuilt from a repository rather than remembered from a bad weekend.

One detail

Troubleshooting is the real curriculum: a GPU quirk, a race condition at startup, a sleep state that fails every third time. Each one teaches more than a machine that "just works" ever could. To be precise — this is a personal configuration, not a distribution; tuned for one user, on purpose.

Stack

Arch Linux · Hyprland · Waybar · Git-managed dotfiles

no public repository yet — ask me about it

ryoku — ws:1 · devarch · hyprland
ws · 1 2 3 4quiet, on purpose
01 — terminal
● 02 — editor
03 — docs
04 — notes
05 — scratch

gaps — 8px · tiling — dwindle · one bar, zero noise

fig. 03 — workspace schematic: gaps, tiling, one quiet bar. A genuine screenshot replaces this once it's worth showing.

More on the shelf

Developer tooling, SAP ABAP deep-dives, small investigations of all kinds — ideas queue up down here and get promoted to an entry above only once they've survived a build. An interest isn't a project until it exists.

02 / About

The common thread

Not a résumé — the pattern that connects an operating system, a market model, and a desktop assistant.

Everything on this site starts the same way: something catches my attention, I can't stop wondering how it actually works, and reading about it isn't enough. So I build a version of it — small, runnable, breakable. Systems under my hands teach faster than systems under a description.

Professionally, I'm the lead engineer of a Raicapp sub-branch that specializes in SAP — and I lead its website development, which in practice means backends, frontends, and everything in between. It's less of a contrast than it sounds: an ERP system is one of the most interlocking machines people ever build, and it behaves exactly like what it is — a system.

The questions don't respect category boundaries, so neither does the work. Some months it's quantitative models and validation design; some months it's a Wayland compositor, a startup trace, or a language I can't leave alone yet. Moving between disciplines isn't a strategy — it's just what following questions looks like in practice.

What I care about most is the reasoning underneath. Getting something to run is a beginning, not a result — the interesting part is why this architecture holds, why that experiment was lying, why the failure only appears under load. A working system is a data point; an understood one is the goal.

“I'd rather understand why something works than stop at the fact that it does.”

03 / Current explorations

Currently circling

Not a skills list — a snapshot of whatever currently has open browser tabs and half-finished notebooks. It moves when the questions move.

  1. 01

    Quantitative ML & time-series validation

    Why most "improvements" evaporate out of sample, and what a validation split that tells the truth looks like.

    experimenting
  2. 02

    Local AI on the Linux desktop

    What small, locally-run models can honestly do offline — and how much of the desktop itself can become the interface.

    building
  3. 03

    System-level debugging & performance

    Startup traces, scheduler behavior, profiling instead of guessing. The machine usually has reasons.

    ongoing
  4. 04

    Frontend architecture & interaction design

    Editorial layouts, restraint as a feature, motion that earns its keep. This site is the practice problem.

    learning
  5. 05

    Mathematics & languages

    The slow-compounding kind of fun. Both punish cramming and quietly reward routine.

    patiently

※ this list gets edited as interests move — it's a snapshot, not a monument.

04 / Contact

Start with a question.

The best way to reach me is to bring something interesting: a project, a half-formed idea, a system behaving badly, a polite disagreement about validation methodology.

I read everything eventually and answer most of it honestly. If you're wondering whether your question is worth writing — it probably is.

The form