Wasted hours, missed deals: Manually cross-referencing dozens of systems daily to explain price shifts and determine next steps.
An AI layer that watches a portfolio all day, scans the market, and gives managers a conclusion they can check before they act on it.
signals rated helpful
helpful signals led to a trade decision
The Cognitive Bottleneck
Every morning, hedge fund managers manually cross-reference six systems — news, X, forums, commentary — to understand what moved their portfolio overnight. That synthesis happens entirely in their head. As one manager put it: "that time is worth a fortune." The question: could an AI layer do that work before they even sit down — without becoming a seventh system to manage?
"The understanding happens entirely in your head — and the time it takes is worth a fortune."
(a portfolio manager, in interview)
The Challenge Was to Integrate Trust into a High-Volume System
1. The volume ruled out a read-and-clear model.
~90 messages per ticker, per day — 3,000+ signals across a book. At that scale, treating signals like notifications to clear one by one wasn't sustainable.
2. Trust is non-negotiable when decisions have financial consequences.
Managers wouldn't act on a signal they couldn't trace to a source and a formula. The system had to show its reasoning, not just its conclusion.
3. Managers already have a filtering logic.
Every interview surfaced the same behavior: managers navigate by entity first — ticker, portfolio, source.
AI Surfaces Signals, the Manager Reads Them, AI Helps Investigate Them
1. Structure: default by priority, narrow on demand
Trade-off: Severity tabs with counts ("40 Critical"), or a single feed ordered by recency and criticality? Chose: A single feed — counts invite clearing, scanning invites reading.
Trade-off: Investigate on the home page, or keep it a glance? Chose: A glance, routed to a dedicated page for investigation. Why: the morning problem is fast synthesis — what changed and why. Investigation is a choice the manager makes after scanning, not the first thing they see.
2. Workflow: scan, read, and ask work as one action
Trade-off: Separate views, or all three on one screen? Chose: One screen, three escalating steps — scan, expand, ask. Why: each step costs slightly more than the last, matching how much the manager has committed to. Verification happens while reading, never on a separate page.
3. Trust: fact, AI interpretation, and checking credibility
Four UX patterns shaped the approach: source attribution; separating fact from interpretation; progressive disclosure, which does two jobs at once — it's the mechanic behind scan/read/ask, and it's what keeps evidence from crowding the conclusion; and human oversight signals, visible cues to confirm before acting.
Shortening the Distance from Signal to Decision
1. Structure
This is where the six-system morning ritual gets replaced: the sources a manager used to check one by one are already scanned, and the news column shows what came out of it — alongside the watchlist tracking price moves on the names they follow. The synthesis now happens on this one page, not across six.
Clicking a signal moves from the "scan and read" state into the full investigation:
2. Workflow
Each card escalates in three taps:
1. Scan (zero tap). Exactly what's needed to judge relevance. (Progressive disclosure.)
2. Read (one tap). The chevron expands the card to reveal the full read.
3. Ask (one more tap). The signal stays in view while it's questioned, and switching to the next one is just as simple. At this volume, the workflow has to support moving on, not just diving in.
3. Trust
signals rated helpful
helpful signals led to a trade decision
Why Interviews Weren't Enough
Feature inflation.
With AI, almost everything is buildable — so nothing was prioritized. Both proactive surfacing and on-demand interrogation shipped, when earning trust on one first might have mattered more.
Interviews revealed wants, not needs.
Accuracy was the real need — assumed, never asked for, because it's a given when reading the news manually.
We never defined a tolerance for AI error.
3 in 10 signals misread their source — a threshold nobody had set.