PhD in Mechanical Engineering — Mathematical Optimization · ML & Perception Engineer @ Doosan Robotics
Mathematical optimization is the through-line — numerical optimization, surrogate modeling & state estimation, carried from theory into 3D perception, sensor fusion, and production systems.
한국어 · English
Three stacks that share a shape — collect → store → research. One Airflow deployment feeds both equity markets; the standalone services sit outside the pipeline.
flowchart TB
subgraph KR ["🇰🇷 Korean equities"]
direction LR
K["kiwoom-client"] --> AF["quant-airflow<br/>DART · KRX · Naver · Toss"] --> DB[("TimescaleDB<br/>delisted included")] --> Q["swing-it"]
NW["krx-news-client"] --> AF
F["krx-fundamentals-client"] --> AF
K --> SC["scalp-it"]
KC["krx-quant-core<br/>shared market rules"] --> SC
KC --> KSIG
KC --> Q
K --> KC
SC -- "ticks · orderbook" --> DB
DB -- "news_judgments<br/>shadow-scored only" --> SC
K --> KSIG["daytrade-it<br/>news-driven day trades"]
NW -- "Toss news, direct" --> KSIG
DB -- "daily bars, read-only" --> KSIG
end
subgraph US ["🇺🇸 US equities"]
direction LR
SH["Sharadar"] --> AFU["quant-airflow<br/>bulk snapshot rebuild"] --> DD[("DuckDB<br/>point-in-time")] --> O["portfolio-research"]
SH --> MS["macro-sector-agent<br/>own PIT DuckDB"]
YF["yfinance"] --> AT["automated-stock-trading-systems"]
end
subgraph CX ["🪙 Crypto"]
direction LR
EX["Exchange APIs"] --> CR["binance-quant-engine"]
end
subgraph SVC ["Standalone services & tools"]
direction TB
FC["fin-checkup"]
end
%% 세로 정렬용 — `~~~` 는 그 자체로 보이지 않는 링크다(linkStyle 불필요).
KR ~~~ US ~~~ CX ~~~ SVC
classDef source fill:#2563EB,stroke:#1E40AF,color:#FFFFFF
classDef move fill:#B45309,stroke:#78350F,color:#FFFFFF
classDef out fill:#059669,stroke:#065F46,color:#FFFFFF
class K,SH,YF,EX,F,NW,FC source
class AF,AFU,DB,DD move
class Q,O,AT,CR,SC,MS,KSIG,KC out
style KR fill:#0F172A08,stroke:#64748B
style US fill:#0F172A08,stroke:#64748B
style CX fill:#0F172A08,stroke:#64748B
style SVC fill:#0F172A08,stroke:#64748B,stroke-dasharray:4 3
Blue — data sources & standalone services · amber — collection & storage · green — research & engines. Dashed — outside the pipeline.
| Project | What it is |
|---|---|
| kiwoom-client |
Kiwoom Securities REST API wrapper — full domestic-equity endpoint coverage & real-time WebSocket feeds · sync + async, auto token refresh · ships an MCP server exposing all 182 REST endpoints plus condition_search as AI-agent tools, real-order calls opt-in only · pip install kiwoom-client |
| quant-airflow |
The one pipeline behind both equity stacks — 16 DAGs. Korea: prices, supply/demand, earnings, consensus, shares outstanding & news/disclosures (via krx-fundamentals-client & krx-news-client) into TimescaleDB over DART · Kiwoom · KRX · Naver · Toss, with delisted-stock backfill so downstream backtests aren't survivorship-biased. Structured LLM judgments over that news/disclosure stream (event type, sentiment, staleness) land in news_judgments — scalp-it scores them in shadow only; none of it reaches an order. US: a daily Sharadar bulk snapshot rebuilt into a DuckDB store and published atomically |
| krx-fundamentals-client |
Korean corporate fundamentals Python client library — financial statements (batched up to 100 tickers/call), valuation metrics, dividends & stock screening (DART + KRX + Naver), no standing server · feeds quant-airflow's earnings/shares/consensus DAGs |
| krx-news-client |
Korean market news & disclosure Python client library — DART filings + Toss Securities, one schema over sources that word the same event differently · feeds quant-airflow's daily_news DAG · pip install krx-news-client |
| krx-quant-core |
The one place the Korean-equity systems agree on market mechanics — KRX tick sizes, upper/lower price limits (checked against real daily bars: the two limits are not symmetric), session times, a dated transaction-tax schedule instead of one hard-coded cost, Kiwoom order guards, DART hard-severity disclosure classification, a kill switch, limit-order fill rules, and the validation statistics (Deflated Sharpe, purged walk-forward, bootstrap). No strategies. scalp-it, daytrade-it and swing-it import it instead of each carrying a copy — every function was ported numerically identical and diffed against the code it replaced before a live trader switched over. Pinned by git tag until the PyPI release |
| fin-checkup |
Risk-disclosure alerts + a financial health checkup over DART & SEC EDGAR — rights offerings, CB issues, audit opinions and delistings, routed to Telegram — collection and classification are verified on real data, the Telegram send itself not yet; 17 statement metrics read as a traffic-light chart against last year, the sector median and the peer percentile. Reports measurements and facts only — never a recommendation |
| swing-it |
KOSPI/KOSDAQ swing research on two axes kept deliberately apart. Observation — sw-flow, a terminal screen of where money went by sector over 5–120 trading days, and sw-ledger, the raw investor-by-sector accounting underneath it; measurements, never a forecast. Judgment — alpha research at the trade-distribution level: walk-forward, random null controls, purged CV, Deflated Sharpe & survivorship-corrected universes, all enforced as CI guardrails. The rejections are the product — pure noise clears “5 of 6 folds positive” 46% of the time, so the test is whether a strategy beats its own randomized version. 5 of 6 alpha hypotheses rejected; PEAD is the one that passed. Formerly kr-quant |
| portfolio-research |
US equity factor engine — point-in-time & survivorship-bias-free, walk-forward optimization gated by Deflated Sharpe & PBO · plus tactical ETF allocation. Ships the rejections too: all 9 pre-registered TAA configs failed the PBO gate, and one headline number was retracted · writeup |
| macro-sector-agent |
Top-down US sector-cycle research pipeline over its own Sharadar-fed point-in-time DuckDB — it asks which industry has been forgotten, not what to buy. The market is cut at a resolution standard sector labels can't show, and an LLM judge argues from evidence over whether a theme is a cycle trough or a structural death, sitting only at that narrow waist — everything above and below is deterministic. The machine never picks — it only excludes; strategy parameters stay out of the repo |
| binance-quant-engine |
Strategy-agnostic Binance USDT-M futures backtest & execution engine — zero lookahead, backtest↔live parity, optional MCP server. Exits are placed as exchange-side algo orders |
| automated-stock-trading-systems |
Backtester for Bensdorp's seven non-correlated trading systems (educational reimplementation) |
Strategies and parameters stay closed. Only structure and discipline are written down. Happy to walk through any of these on request.
| Project | What it is |
|---|---|
| scalp-it |
Korean intraday strategy validation framework + live tick/orderbook collection + a live execution loop — since late August 2026 its detector sends real orders under a hard-capped order size, a price band, a daily order cap and a consecutive-loss kill switch. Ticks cannot be backfilled, so a missed day is gone for good. Pre-register, measure once. No re-tuning to revive a rejected hypothesis |
| quantbox |
Binance USDT-M futures breakout/momentum system — VR compression squeeze + MA cluster squeeze. Traded live; the live bot is currently paused. binance-quant-engine is the public extract with the strategies removed |
| momentum |
US equity screener — Minervini Trend Template + VCP pattern, DuckDB-cached, CLI-driven |
| daytrade-it |
Korean-equity (KOSPI/KOSDAQ) news day-trading system, the second live trader on trader — unattended real orders from 2026-09-14. A daemon polls Toss news, has Claude extract facts only — is the company the article's subject, is it new, which direction — never a trade call; scoring is code. A BUY needs the stock not to have already run, the same no-chase finding scalp-it reached on its own. That rule was picked on an outcome-labeled eval with a held-out split, over a price-only ML model (38.5% vs. a 38.1% baseline) and over asking the model to predict the reaction, which did worse — and the held-out split is only 10 trading days. Execution: 1 share, LIMIT, price band, one entry per ticker per day, at most 2 open, unfilled entries cancelled, and from 15:00 it sells back only the shares it bought, pricing each retry further below the bid through the closing auction. Full-text articles are scored in shadow to collect forward evidence. Entries are refused while scalp-it's DART cache shows a hard-severity filing for the ticker in the past week — delisting risk, embezzlement, rehabilitation — and also while that cache is missing or stale, so the gate can't pass trades silently. Redeveloped from gpt-quant-v2 |
| crypto-pair-trading |
First iteration of the crypto pair-trading framework — predecessor of quantbox |
| resume-private |
Private résumé source (LaTeX) |
Most of these aren't just repos — they're running right now, across two hosts that share the same repos but not the same job. One rule draws the line: what can't be redone stays on trader; what can be rerun lives on simnode. Ticks and orderbook snapshots can't be backfilled, so a missed market hour is gone for good — a failed batch is just rerun tomorrow.
flowchart TB
KW(["Kiwoom<br/>one app key"])
subgraph T ["🖥️ trader — live"]
SC["scalp-it<br/>collector + live orders"]
DT["daytrade-it<br/>news → 1-share trades"]
DD[("dart.db<br/>local SQLite")]
RP[("TimescaleDB<br/>standby replica<br/>no app reads it")]
end
subgraph S ["🖥️ simnode — reproducible"]
AF["quant-airflow<br/>16 DAGs"]
PR[("TimescaleDB PRIMARY<br/>kr_quant · gptquant")]
RS["swing-it · portfolio-research<br/>macro-sector-agent · momentum<br/>backtests"]
end
KW <-->|"ticks · orderbook · orders"| SC
KW <-->|"quotes · orders"| DT
DD -->|"DART every 10 min"| SC
SC <-->|"writes ticks · orderbook, spooled<br/>reads universe · regime"| PR
DT <-->|"writes signals<br/>reads daily bars"| PR
AF -->|"prices · news_judgments"| PR
PR -->|"streaming replication"| RP
PR --> RS
classDef live fill:#B45309,stroke:#78350F,color:#FFFFFF
classDef repro fill:#059669,stroke:#065F46,color:#FFFFFF
classDef store fill:#2563EB,stroke:#1E40AF,color:#FFFFFF
classDef ext fill:#64748B,stroke:#334155,color:#FFFFFF
class SC,DT live
class AF,RS repro
class RP,PR,DD store
class KW ext
style T fill:#0F172A08,stroke:#64748B
style S fill:#0F172A08,stroke:#64748B
Every arrow into or out of a database crosses the LAN to simnode — nothing on trader reads its own replica. The replica is there to be promoted, not queried.
trader — the live box |
simnode — the research box |
|
|---|---|---|
| Job | Irreversible, wall-clock bound — market hours happen once | Reproducible — orchestration, batches, research |
| Runs | scalp-it tick/orderbook collection and live orders · daytrade-it news-driven live orders · quantbox (paused) · kiwoom-client development, because a broker session is one-per-key and it lives here |
Airflow scheduler & webserver (16 DAGs) · TimescaleDB PRIMARY · swing-it, portfolio-research, macro-sector-agent, momentum · post-close research batches · backtests |
| Shared repos | quant-airflow exists here only as a git sparse-checkout — the replica's compose file, the schema, and the one .env every live process sources for broker keys, the Claude key and the DB DSN. Research-only repos aren't here at all |
The canonical full clones |
| TimescaleDB | Standby streaming replica | PRIMARY — every read and write, from both hosts |
Two live systems, one broker account. They trade on different evidence and neither knows the other exists — the account is the only thing they share.
flowchart LR
subgraph SCP ["scalp-it — fully unattended, 08:55–15:20"]
WS["Kiwoom WebSocket<br/>ticks · orderbook"] --> DET["pair detector<br/>leader → follower"]
DART[("dart.db")] -->|"breaking disclosures"| DET
DET --> GRD["order guard<br/>size cap · price band · daily cap<br/>2-loss kill · pair_STOP"]
end
subgraph DTP ["daytrade-it — fully unattended, 09:00–15:30"]
TOSS["Toss news"] --> CL["Claude<br/>facts only, no trade call"]
CL --> TC["no-chase filter<br/>vs. prior close"]
TC -->|"BUY only"| SIG[("gptquant<br/>trading_signals")]
SIG --> AT["AutoTrader<br/>1 entry per ticker · ≤2 open<br/>cancel unfilled · sell own shares 15:00–15:30"]
AT --> ET["execute_trade<br/>1 share · LIMIT · price band<br/>DART gate · live_STOP"]
end
ACC{{"Kiwoom account<br/>same app key"}}
DART -->|"hard filings, 7 days<br/>stale → refuse"| ET
GRD -->|"real order"| ACC
ET -->|"real order"| ACC
ACC -. "holdings → refuse to buy a held ticker" .-> GRD
ACC -. "holdings → refuse a second entry" .-> ET
classDef live fill:#B45309,stroke:#78350F,color:#FFFFFF
classDef store fill:#2563EB,stroke:#1E40AF,color:#FFFFFF
classDef gate fill:#B91C1C,stroke:#7F1D1D,color:#FFFFFF
classDef ext fill:#64748B,stroke:#334155,color:#FFFFFF
class DET,CL,TC,AT live
class DART,SIG store
class GRD,ET gate
class WS,TOSS,ACC ext
style SCP fill:#0F172A08,stroke:#64748B
style DTP fill:#0F172A08,stroke:#64748B
Red — the last check before an order. Dashed — what each side reads back from the account before it buys.
A weekday, in KST:
| Time | trader |
simnode |
|---|---|---|
| 08:30–08:45 | Morning report on yesterday · DART refresh into dart.db · two read-only pre-open checks, one per live system |
premarket_news_judgment — Toss + DART → Claude → news_judgments |
| 08:55 | Both launchers start — scalp-it's detector reads today's universe from the PRIMARY; daytrade-it snapshots the account's holdings so it only ever sells what it bought | |
| 09:00–15:20 | scalp-it collects and trades · daytrade-it enters 09:00:30–14:30 and flattens its own shares from 15:00 · DART every 10 min · tick health at 09:10 and 10:00 | daily_news and a collection catch-up at 10:05 · Airflow health check at 11:35 |
| 15:20–16:10 | scalp-it stops at 15:20; daytrade-it stops reading news but manages exits through the 15:30 closing auction (15:40 kill as a backstop) · same-day morning report · tick sanity · theme snapshot → PRIMARY | 16:00 daily_collection, daily_earnings · 16:05 daily_news again · news-judgment shadow report |
| 16:55–19:00 | Price adjustment · consensus · Sharadar (Tue–Sat) · swing-it daily report · coverage · Google Drive backup |
- The same repo on both hosts doesn't make both real. Config edits are made in
simnode's full clone and pushed;traderonly pulls. A sparse checkout makes that hard to get backwards. - Which host is PRIMARY is state, not code —
pg_is_in_recovery()answers it, and the answer has already flipped once. The primary started ontraderso the collector couldn't be killed by a LAN blip; it moved tosimnodethe day the collector grew reconnect + disk spooling and proved itself in production. The demoted host was rebuilt as the replica, compose file name and all. - The live box leans on the research box every morning. Spooling protects ticks going out; it does nothing for inputs coming in. scalp-it's universe and market regime, and daytrade-it's prior close, are all read from the PRIMARY — built from yesterday's 16:00 collection. If that read fails, scalp-it falls back to a fixed pair list and daytrade-it enters nothing.
- Two systems, one account, no shared kill switch. Both ask the broker for holdings before buying and refuse a ticker the account already holds, so whichever enters first owns that ticker until it's closed. They also draw on the same cash, which is why daytrade-it holds at most two positions. And daytrade-it leans on scalp-it: its disclosure gate reads scalp-it's DART cache, so if that cron stops, daytrade-it stops buying. But
pair_STOPstops only scalp-it, andlive_STOPstops only daytrade-it's entries — its exits keep running, because a kill switch that strands an open position isn't a safety feature. - Backtests run on
simnode, full stop.trader's CPU belongs to the live daemons, so the backtest entry points of both live systems check the hostname and refuse to run anywhere else — one guard inkrx-quant-core, not a copy per repo.swing-itisn't checked out ontraderat all anymore. - Three systems, one set of market rules. The tick table, price limits, order-guard checks and disclosure taxonomy used to live in each repo, and had already drifted — the same
normalize_codepasted twice, cost assumptions scattered as0.0023,0.0034and0.0064with no notion that the tax changed four times in four years. They now come fromkrx-quant-core. Moving the live traders onto it was done the way you'd move money: side-by-side diffs against the old code on randomized inputs (tens of thousands of orders, every filing in the DART cache), the live checkout untouched until the new one matched, then both pre-open checks run end to end. - Moving a batch didn't move its clock. Cron triggers are anchored to when the data is final in the DB, not to the host, so the schedule read identically before and after the split.
- Backups follow the primary. After the switch both hosts still ran the Drive backup at 19:00 into date-named files, so whichever finished later won — and that could be
trader's stale standby dump. It now runs onsimnodeonly. - The one thing that had to be switched off: a health guard that restarts a downed DB container. Left running, it would have resurrected the demoted primary into a split brain. The script stays on disk for the day
traderhosts the primary again.