Platform

Written and run in-house

Market data capture, the historical archive, backtesting, execution and reconciliation are all written by Corey Olson. Each decision is logged with its timestamped inputs and the rule that fired.

Market data
~20k messages a second, 62k at peak
Archive
50+ TB
Replay
A trading day in ~10 seconds
Fleet
15 machines, 3 SANs
01Components

The main pieces

  1. Market data
  2. Archive
  3. Strategy definitions
  4. Backtesting
  5. Replay ladder
  6. Execution
  7. Reconciliation

Market data

Equity, index and options data, captured and normalized in real time.

Archive

Market history, including option quotes, as it was known at the time.

Strategy definitions

Each strategy is defined once and runs unchanged in research, paper and live.

Backtesting

Full option chains, replayed in parallel across machines.

Execution and reconciliation

Order placement and lifecycle, with fills reconciled against the broker.

Monitoring

Services log to a central event journal, run continuous health checks and can be stopped by hand.

02Replay ladder

Six test modes before live

Each mode swaps a research stand-in for production code. Live is the reference: a faster mode is trusted only once it matches the one above it, so when a backtest and live results disagree, the first mode where the gap appears points to the layer at fault.

The modes span four orders of magnitude in speed, from real time down to sub-second cached reuse. A normal trading day replays in about ten seconds.

See the systems running

Each mode adds one production layerReplay ladder
  1. Mode 1 DiscoveryFast, broad evaluation over point-in-time history
  2. Mode 2 ScreeningFull strategy rules, fast enough for broad research
  3. Mode 3 IntegrationProduction code replaces research stand-ins
  4. Mode 4 ReplayProduction order handling, fills modeled from archived quotes
  5. Mode 5 SystemThe full production stack, run on history
  6. Mode 6 SimulationReal-time pace, monitored the same way as live
  7. Live ProductionThe live market and the broker
03AI

Where the AI sits

Agents run as supervised, long-lived services. Work lands on a queue and a free agent claims it. For model research, an LLM proposes structural changes and a constraint engine accepts or rejects each one on evidence.

  • Eight statistical gates stand between a proposal and acceptance.
  • The LLM searches only inside what the engine allows; the problem and the deploy call stay with a person.
  • Every deploy passes eleven blocking checks in a CI pipeline written in-house.
Who decides whatAuthority
  1. Human Sets the problem and makes the deploy call
  2. Engine Decides what is admissible; eight statistical gates accept or reject each change
  3. LLM Proposes changes, searching only inside what the engine allows