AI-Assisted Trade Navigation

Context and timeline

In Q1-Q2 2022, we tested whether free-text analyst requests could safely drive navigation inside a trade analytics product. The feature was approved in April 2022 and deployed publicly soon after.

Problem

Analysts had to move through several filters and views to compare symbols, time ranges, traders, and P&L. The data model was structured, but the navigation was slower than the work required.

Solution

We added a small prompt box that translated analyst intent into validated navigation parameters. Users could type a request, review the parsed output, and jump directly to the right view without rebuilding filters by hand.

Key design choices

• Assist, not replace: the feature sped up navigation but never removed manual controls.
• Validation first: parsed output was checked against allowed symbols, intervals, and target views before execution.
• Clear feedback: users could review and edit the parameters before applying them.
• Limited scope: the model handled navigation only, not analysis, prediction, or advice.

Technical approach

• Normalized trade data feeding an analytics API
• Constrained prompt template with known domain terms
• GPT-3 completion call returning compact navigation hints
• Parser and validator mapping the result into internal parameters
• Execution and routing layer moving users to the correct view
• Logging of intent, parsed parameters, and latency for review

Outcome

• Fewer steps for common exploration tasks
• Faster movement across trades, traders, positions, and P&L views
• Less abandonment of partially configured queries
• A clear example of AI helping inside a narrow, well-bounded workflow

Responsible AI safeguards

• Scope restriction: the model only helped with navigation parameters.
• Hard validation: unsupported values were rejected or sent back for confirmation.
• Audit trail: anonymized prompts and outputs could be reviewed for quality.
• User control: the final action always stayed with the user.

Why it matters

This was not a chatbot demo. It was a small, useful AI layer inside a real product, designed to reduce friction without giving up control, validation, or observability.

Related services

If you want a similar assistive layer in your product or internal workflow, see our Data & Analytics, Architecture & Design, and Custom Development services, or start with the Backlog to First Release program.