Execution

Technical sheet · Programme insert

Live trading execution

Private system · Strategy, universe, positions and P&L stay private

A private live-trading system on AWS that scans intraday, filters by factor, and routes bracket orders through Alpaca into a SQLite execution ledger. This sheet prints the order path, the invariants that keep it safe, and one measured regression with its method. Every number here carries its method line.

Role: Co-engineer · since Apr 2026
Broker: Alpaca · bracket orders
Market data: Massive
Live store: SQLite · one writer
Warehouse: S3 Parquet
Publisher: AWS Lambda → CloudFront / S3

Featuring fill-dedupe keys · startup reconcile · bracket / OCO · flag-gated routing · after-close watchdog · SSM-only deploys · content-addressed snapshots

01System

Order path left to right
One host · one writer

Eight stages · what each one is responsible for
Stage	Responsibility	Notes
Market data	Intraday feed into the scanner.	Massive
Scanner	Batches the universe each cycle. Event-based waits instead of polling; warm-cache fast paths; lazy-load startup.	The scan batch is the unit measured in 03.
Factor filter	Scores candidates. Reject reasons are vectorized so every rejection is explainable after the fact.	Factors private.
Broker	Bracket orders: entry plus exit legs. Live routing moved onto Alpaca behind a flag.	Alpaca · IBKR route retired
Execution ledger	Orders and fills, deduped by fill key; reconciled against the broker at startup.	SQLite · the subject of 03
Warehouse	Durable history for research.	S3 Parquet
Publisher	Renders the dashboard as a pure function of content-addressed input snapshots; compare-and-set supersede; skip-unchanged publishing.	AWS Lambda
Dashboard	Public, immutable render with redacted live logs.	CloudFront + S3 (OAC)

SQLite keeps the live state on one host as one file: one writer, no network hop inside the order path, and a single artifact to reconcile against the broker at startup. The regression in 03 is the cost of that choice the moment a second writer appears.

Era diagrams for this path, Fig.01 through Fig.06, are printed on the architecture hub.

02Execution safety

Mechanisms as named in the resume
Test and runbook evidence is private

Invariant: A fill is applied to the ledger exactly once.
Mechanism: Fill-dedupe keys on the shared execution ledger.
Status: Live
Invariant: Broker and ledger agree before the first order.
Mechanism: Startup reconcile of broker state against the database.
Status: Live
Invariant: Every entry carries its exit.
Mechanism: Bracket orders with OCO legs; bracket / OCO fixes landed before the Alpaca cutover.
Status: Live
Invariant: A routing change is reversible.
Mechanism: Live routing migrated onto Alpaca behind a flag; the IB Gateway path was removed only after cutover.
Status: Live · IBKR is the retired path
Invariant: Nothing runs unattended after the close.
Mechanism: After-close watchdog.
Status: Live
Invariant: No secret lives on the host or in the repo.
Mechanism: Live secrets in SSM Parameter Store; SSM-only deploys.
Status: Live
Invariant: The public dashboard cannot leak or lie.
Mechanism: Pure function of content-addressed snapshots; compare-and-set supersede; skip-unchanged publishing; redacted live logs; CloudFront / S3 origin access control.
Status: Live
Invariant: Only one process writes the live store.
Mechanism: SQLite single writer. Broken on 2026-10-08 by a backfill cron on the same host; restored by removing the cron on 2026-10-09. See 03.
Status: Restored 2026-10-09

03Measured

Values marked ~ are read from logs, not a controlled benchmark
Fix applied 2026-10-09 · cron removed from the live host

Scan batch latency · before and after the write-lock fix
Metric	Baseline	2026-10-08 regression	2026-10-09 post-fix
Scan batch latency, p95	~2.4 s	~9.2 sBackfill cron active	≈1.55 sCron removed 2026-10-09 · measured 2026-10-12 13:26–14:59 ET · p50 0.71 s, p99 1.69 s, n=276 · morning not measured

Ink points are log reads marked ~ · the vermilion rule marks the 2026-10-09 fix (cron removed)

Method

Measure: Scan batch latency: wall-clock from the start of a batch to the last symbol scored, taken from the scanner's own batch timings.
Window: Before: 2026-10-08, full regular session, 09:30–16:00 ET. After: 2026-10-12, afternoon only, 13:26–14:59 ET (morning not measured). p50 and p95 over every batch in the window.
Host: The live scanner host, unchanged between before and after.
Before: Session of 2026-10-08, backfill cron active (every 5 minutes).
After: 2026-10-12, cron removed; afternoon window only (13:26–14:59 ET, n=276 batches). Morning not measured.
Publish: Values land in the table with their session date. A tilde marks a read from logs rather than a controlled benchmark.

On 2026-10-08 scan-batch p95 rose from about 2.4 s to about 9.2 s with no change to the scanner itself. The cause was a backfill cron running every five minutes on the same host. SQLite admits one writer at a time; while the backfill held the write lock, the scanner's own ledger writes queued behind it, so each batch inherited the backfill's lock-hold time.

The cron was removed from the live host on 2026-10-09. The single-writer invariant was restored. On the 2026-10-12 afternoon (13:26–14:59 ET, n=276) p95 measured ≈1.55 s (p50 0.71 s, p99 1.69 s); the morning was not measured. The 2026-10-08 baseline was 9.2 s.

04Boundary

What this sheet shows and what it does not

Public

This sheet: the order path, the invariants in 02, and the method in 03.
The architecture hub: era diagrams Fig.01–Fig.06.
The live research dashboard ↗, published by the Lambda in 01.
The reel on the full programme and in the resume.

Private

Strategy, factors, and the symbol universe.
Positions, orders, and fills.
P&L. The dashboard shows its own daily figure; it is not reproduced on this sheet.
Source code, tests, and runbooks.

Ink · oxblood · one vermilion spot on warm stockPrinted 2026-10-09 · Cron removed 2026-10-09yplmicro@gmail.com