01 Home02 Research03 Projects04 Notes05 About06 Lab
Home/Projects

Systems, evaluations
and useful failures.

What I build and test on my own: two projects with case pages, and the daily routines that keep my inputs current. Academic research is on the Research page.

Independent system · 2026 · ongoing

Helix

I designed and built a research and paper-trading system in which a signal must clear provenance, backtest, risk and validation checks before it reaches a simulated trading loop, across US, China A-share and Hong Kong equities.

Its own deflated-Sharpe gate rejects the book shown here, and no absolute backtest returns are quoted: survivorship inflates them by about five points a year.

Read the Helix case →
Helix · Research terminal · snapshot 27 Jul 2026
The Helix research terminal: a four-strategy book with its combined equity curve, signal-parity and risk checks passing, and the deflated-Sharpe validation failing.
Helix · Exhibition
The Helix exhibition page: 23 live browser experiments in five wings.
Validation gate · this snapshot
Signal parity · passRisk · passDeflated Sharpe 0.21 · fail
The system rejects its own book.

Real screenshots of the Helix web front. The terminal shows a saved snapshot (27 Jul 2026, 50 liquid US stocks): an in-sample backtest whose combined book fails Helix’s own deflated-Sharpe gate. Its numbers are not performance claims.

Independent evaluation · Jul – Sep 2026

Kronos under the microscope

I tested the public Kronos-small checkpoint, a financial foundation model by its original authors, against simple baselines on CSI300, and added a bounded KV cache that made its CPU inference 11.7× faster with bit-identical output.

+1.93% a year after costs against +1.66% for simple reversal, in one universe and one window; the result moves with the number of forecast paths.

Read the Kronos case →
eval/topk_backtest.py · CSI300 · Kronos-small
Cumulative excess return over CSI300 after costs: Kronos-small ends at plus 1.3 percent, reversal plus 0.7, random minus 2.6, momentum minus 16.3.
Bounded KV cache11.7× fasterBit-identical output · CPU, 400-bar lookback
Ranking signalp = 0.295Not significant (Newey–West, 5-path run)

Redrawn from the evaluation’s daily reports: compounded excess over CSI300 after costs. The annualized mean, +1.93%, is the figure quoted elsewhere.

Daily routines.

Scheduled agents · designed by me

Four scheduled Claude Code agents keep my inputs current. I designed each one: its sources, its format and its checks. The agent runs it on schedule and files the output. They are working tools, not published research.

flashcards.html · Daily Quant Interview
The Daily Quant Interview flashcard app: a grid of 552 problems and one open problem on the bias of the uniform maximum-likelihood estimator.
Weekdays · 09:00 SGT · since Jun 2026

Daily quant interview

Eight interview problems each weekday across maths and statistics, algorithms, brain teasers and quantitative finance, with a timed mock on Fridays. A concept ledger blocks repeats, and anything I mark as missed comes back later as a fresh re-test.

Checks
Coding and numerical solutions are run in Python before archiving; sources are tagged, or marked self-generated.
Output
552 problems by 24 Sep 2026, as a flashcard app and Obsidian notes.

Solutions are drafted by the agent; the Python checks cover only the parts that can be computed.

Investor Compass · The Board
Investor Compass: the board of disclosed moves by Buffett, Burry and Ackman, each with a why and a critical lens.
Daily · 08:00 SGT · since Jul 2026

Investor Compass

A static site for learning from eight public investors’ reasoning rather than copying their positions. Each morning the agent checks new filings, ARK’s daily trades and public posts, and records only genuinely new activity; quiet days stay quiet.

Checks
Every entry needs a why and a critical lens, the reason copying it could be a mistake. The log is append-only, and a browser test must pass before each commit.
Output
187 observations across 8 investors by 24 Sep 2026.

13F filings arrive about 45 days late and cover long US equities only. A learning tool, not investment advice.

GitHub Intelligence · research desk
The GitHub intelligence dashboard: 1,038 repositories, weekly top gainers and a scatter of adoption against growth.
Daily · 09:00 SGT · since Jun 2026

GitHub daily highlight

A morning brief on repositories and agent skills worth knowing in AI agents, data science and quantitative finance, drawn from GitHub Trending and Trendshift, with a weekly view every day and a monthly view on Mondays.

Checks
Star counts come from the GitHub API on every run and are kept as daily snapshots, so growth is measured rather than estimated.
Output
1,038 repositories tracked, 70 daily snapshots from 22 Jun to 23 Sep 2026.

Stars measure attention, not quality.

US Market Desk · 10 Sep 2026 session
A US close brief: the verdict, the close tape and the week’s regime shift.
Asia Close Wire · 10 Sep 2026
An Asia close brief: the headline, the regional tone gauge and the overnight lead.
Daily · 06:00 and 09:00 SGT · US since Jun 2026 · Asia since Aug 2026

US and Asia close briefs

One brief per completed session: macro, indices and mood, leaders and guidance, a watchlist, a judgement and what to watch next. Every data point carries its reason, its influence and a one-week to six-month lookback, with an interactive page beside the text.

Checks
Lookbacks are anchored to the archive’s own recorded closes rather than a fresh search, and unverified figures are marked as such.
Output
Each run files its brief as a pull request to a private archive.

Informational only, not investment advice.