TradersGPT

Traders need an answer that pulls from web search, financial document retrieval, and live market data from external APIs — and they need it to be right often enough to rely on.

Our agents ran sequentially and took over 200 seconds to fetch, analyze, and visualize. The team treated that as the cost of doing the work properly.

I thought three minutes was a product failure rather than a performance one. Nobody waits three minutes for an answer they can get worse and faster elsewhere.

I rewrote the architecture overnight and pitched it at Monday standup: parallel agents, query routing, contextual visualization, and SSE streaming from the backend. Latency dropped to around 30 seconds including thinking time, and because results streamed as they arrived it felt instantaneous. It has been in production since.

Tool selection at 120+ tools. Scale breaks routing before it breaks anything else, so most of the design effort went into the layer that decides what to call, not the tools themselves.

I built the orchestration, the retrieval, the visualization layer, and the evaluation harness — LangSmith tracing across agent handoffs, tool calls, and orchestration paths — because without it you cannot tell whether a change made the system worse.

It held uptime that traders were willing to rely on, which is a higher bar than a benchmark. I designed the architecture and implemented it end to end.