All work

Strategy Copilot

A team-built LLM copilot that turns plain English into backtestable strategies. My part: grounding, design system, save.

Context
Built at a crypto derivatives exchange
Year
2026
Role
Contributor to a team-built system
Stack
Python, FastAPI, Claude tool use, React

The problem

The copilot drafts trading strategies from plain English and stages backtests on the engine. Teammates built its chat loop, tools, guardrails and eval gate. When the engine gained gold options, the copilot still had BTC and ETH facts written into its prompts, and its web app broke down at common screen sizes.

  • Facts about what the engine supports come from the engine, never from the prompt.
  • Older engines without the new config must keep working.

Architecture

Fig. 1 Each chat turn passes budget and guardrail checks, then a tool loop composes the reply. A number check rejects any figure no tool returned and sends the reply back to be regenerated, before the turn is saved and streamed.

What I built

  1. 01Grounded per-asset expiries, settlement times and backtest eligibility in the engine's live config, so the copilot could offer XAUT options without hard-coding them, with a cached background refresh and a fallback.
  2. 02A design-token system for the web app, enforced by a dependency-free lint, with layout modes that keep the chat readable at every breakpoint.
  3. 03Save from backtest results: keep the exact tested strategy in one click, even if the trader switches chats mid-save.

By the numbers

UTC gold settlement, read from engine config
16:00
end-to-end tests on the redesign
48/48