Taurex is a regulated multi-asset trading platform. The company operates in a crowded brokerage market where competitors primarily differentiate on spreads, execution speed, and platform compatibility.
While some competitors have begun introducing AI-adjacent features (typically third-party behavioral analytics integrations or lightweight trading record analyzers), none have built a purpose-designed data engineering pipeline that produces auditable, pre-computed coaching from the platform’s own trade data.
Taurex saw this as an opportunity to strengthen trader engagement, improve client outcomes, and differentiate its mobile app through an AI-powered Trading Coach that delivered personalized trading insights, behavioral feedback, and coaching tips.
The initial iteration of the system took three minutes to return AI reports. Load testing showed it could support 25–30 concurrent fresh trader accounts with nearly 100% success, but success dropped to almost 60% at 50 concurrent users.
Taurex brought in Taazaa to improve the AI Trading Coach. The goal: Support 500 requests and reduce response time to 90 seconds.
Taazaa built the Taurex Trading Coach as a daily analytics pipeline with a strict separation of concerns. Raw trade data is pulled from the platform’s MSSQL database into PostgreSQL via a daily ETL process. Twelve dbt metric models then compute win rates, drawdown patterns, risk exposure, position sizing consistency, behavioral indicators, and other performance metrics.
These pre-computed analytics produce a composite Trading Health Score on a 0–100 scale. The score and its component metrics are then fed into an LLM to generate personalized, mobile-optimized coaching reports. The pipeline is orchestrated via SQS and deployed on ECS Fargate.
One design principle was critical: the LLM performs no mathematical computations. Every number, every ratio, and every trend in the coaching report are computed deterministically in the dbt layer. The LLM’s role is strictly narrative. It translates structured analytics into plain-language coaching that a retail trader can act on.
When a trader sees a Health Score of 62 with a note about inconsistent position sizing, every figure in that report is auditable and reproducible. The LLM cannot hallucinate the math because it never performs any.
Taazaa redesigned the Taurex AI Trading Coach’s architecture, moving from synchronous compute-on-read to pre-compute-and-serve. The new structure drastically improved the application’s performance.
Average request latency dropped from 148 seconds to 22 seconds, and report retrieval load testing showed approximately 70 requests per second without degradation. Token usage dropped from roughly 50,000 per request to 6,000–8,000.
A later API load test ran 100 virtual users for 30 minutes, with 75,523 successful requests, no failures, an average of 42 requests per second, and an average response time of 383.42 milliseconds. 95% of requests were completed in less than 650 milliseconds.
The 20-user comparison improved average report delivery time from 1.5 minutes to less than half a second.
To date, no direct competitor in the retail brokerage space has replicated this architecture.
