01
Bespoke AI
Bespoke came to me with a problem every investment desk knows but nobody had solved cleanly: too much data, too little signal. They were building the first knowledge management system purpose-built for asset managers, & needed a product that felt as intelligent as the infrastructure underneath it.
Role:
Principal Product Designer
Industry:
Fintech / Asset Management
Duration:
20 weeks
Challenges
Complexity without the black box. The architecture, a RAG-powered Intelligence Vault, an agentic router, and a proactive signal engine, was technically defined. The real problem was translating that system into something PMs and analysts would trust on a live trade.
Skeptical, expert users. Financial professionals don't forgive vague interfaces. Every design decision had to make the system's reasoning visible without adding friction to the decision itself.
Signal, not noise. Proactive alerts are only useful if they're relevant. A system that surfaces everything surfaces nothing. Getting the information architecture right meant understanding the difference between "what happened" and "what happened to you."
My Approach
Discovery Phase
Workflow mapping: Spent time with the founding team, former traders and PMs, tracing the actual shape of a research day from first signal to final decision.
Key insight: The bottleneck wasn't finding information. It was knowing which information to act on. That distinction shaped everything downstream.
Design and Development
Reframe from completeness to confidence: Early concepts surfaced everything the Vault knew in response to a query. Comprehensive, but overwhelming. The shift came when we stopped designing for completeness and started designing for confidence.
Router as UX centerpiece: The Agentic Router became the defining interaction; queries needed to feel answered, not just returned. The response pattern showed source, reasoning, and action in one view.
Tradeoffs: Showing less required more decisions about hierarchy. Every field on screen had to justify its presence against a user mid-decision.
Testing and Iteration
Newsfeed failure: The proactive alert surface broke in testing, not technically but behaviorally. Users ignored alerts that felt generic.
The fix: Reframing from "what happened in the market" to "what happened that affects your book" doubled engagement. Signal relevance, not volume, was the real design problem.
Results
Bespoke launched with a system investment managers could query in plain language and trust the answer. Early users reported cutting research prep time significantly. The Newsfeed became the stickiest surface, the thing PMs checked before morning meetings. The product carved a clear position in a space most fintech incumbents hadn't touched.
Future Plans
A mobile-first alert layer for on-the-go PMs who need signal without a full query workflow.
A compliance overlay that surfaces mandate violations before a trade executes, embedded directly in the decision flow rather than as a separate review step.
Conclusion
Bespoke is a bet that the investment desk of the next decade runs on institutional intelligence, not just market data. Building the product meant learning to design for trust under pressure, where the cost of a confusing interface isn't frustration but a bad trade.
Challenges
Complexity without the black box. The architecture, a RAG-powered Intelligence Vault, an agentic router, and a proactive signal engine, was technically defined. The real problem was translating that system into something PMs and analysts would trust on a live trade.
Skeptical, expert users. Financial professionals don't forgive vague interfaces. Every design decision had to make the system's reasoning visible without adding friction to the decision itself.
Signal, not noise. Proactive alerts are only useful if they're relevant. A system that surfaces everything surfaces nothing. Getting the information architecture right meant understanding the difference between "what happened" and "what happened to you."
My Approach
Discovery Phase
Workflow mapping: Spent time with the founding team, former traders and PMs, tracing the actual shape of a research day from first signal to final decision.
Key insight: The bottleneck wasn't finding information. It was knowing which information to act on. That distinction shaped everything downstream.
Design and Development
Reframe from completeness to confidence: Early concepts surfaced everything the Vault knew in response to a query. Comprehensive, but overwhelming. The shift came when we stopped designing for completeness and started designing for confidence.
Router as UX centerpiece: The Agentic Router became the defining interaction; queries needed to feel answered, not just returned. The response pattern showed source, reasoning, and action in one view.
Tradeoffs: Showing less required more decisions about hierarchy. Every field on screen had to justify its presence against a user mid-decision.
Testing and Iteration
Newsfeed failure: The proactive alert surface broke in testing, not technically but behaviorally. Users ignored alerts that felt generic.
The fix: Reframing from "what happened in the market" to "what happened that affects your book" doubled engagement. Signal relevance, not volume, was the real design problem.
Results
Bespoke launched with a system investment managers could query in plain language and trust the answer. Early users reported cutting research prep time significantly. The Newsfeed became the stickiest surface, the thing PMs checked before morning meetings. The product carved a clear position in a space most fintech incumbents hadn't touched.
Future Plans
A mobile-first alert layer for on-the-go PMs who need signal without a full query workflow.
A compliance overlay that surfaces mandate violations before a trade executes, embedded directly in the decision flow rather than as a separate review step.
Conclusion
Bespoke is a bet that the investment desk of the next decade runs on institutional intelligence, not just market data. Building the product meant learning to design for trust under pressure, where the cost of a confusing interface isn't frustration but a bad trade.