Back to projects

In progress

This page describes the intended system and the work ahead. Performance results will be added after experiments are run.

In Progress2026

Trading Infrastructure Optimizer

An electronic trading testbed for measuring order flow, tail latency, and failure recovery. The longer-term goal is an optimizer that tests small configuration changes and keeps only measured improvements.

The question

Can infrastructure tune itself without losing correctness?

The project starts with a simplified exchange: trading clients send orders through a gateway to a matching engine, which produces market data and execution reports. Instrumentation will show where time is spent before any optimizer changes the configuration.

This is an in-progress system design. The architecture and experiments below are planned work, not completed benchmark results.

Planned order path and control loop.

01

Trading client

Generates orders

02

Order gateway

Validates and routes

03

Matching engine

Executes against a book

04

Market data

Publishes book and trades

Observe latency and failures → propose a bounded change → run an experiment → keep or roll back

The first milestone is a working order path and baseline telemetry. The optimizer comes after there are measurements to compare.

How would an experiment be judged?

Choose a planned stress scenario to see the signals the control loop would inspect. These are experiment designs, not results.

Measure
Queue depth, p99 order latency, throughput
Decision rule
Compare bounded batching and polling changes; revert if p99 or correctness worsens.

Tools under consideration

Aeron (planned)Java / C++PythonGCP (planned)

Project plan

  • —The proposed testbed has trading clients, an order gateway, matching engines, a market-data service, and a separate execution-report path.
  • —The first milestone is a working order path with latency, throughput, and queue-depth instrumentation before introducing automated tuning.
  • —Later experiments will compare baseline and tuned configurations under burst traffic and failures, with checks that can reject or roll back harmful changes.