Role
Role
Senior Product Designer
Senior Product Designer
Sr. Product Designer
Team
Team
2 Engineers
Frontend Engineer
2 Engineers
Duration
Duration
~8 months · 2025-2026
~8 months · 2025-2026
Status
Status
Shipped · Live
AI Chat
AI Chat
An AI assistant integrated into a modular trading platform — designed to analyse charts, assist with risk management, answer market questions, and generate indicator code.
An AI assistant integrated into a modular trading platform — designed to analyse charts, assist with risk management, answer market questions, and generate indicator code.
Live product: takeprofit.com
Product Design
Product Design
AI / LLM
Web
Web
Lead
Lead
01. A working AI prototype became a full-scale product feature
01. A working AI prototype became a full-scale product feature
TakeProfit is a modular trading interface. In fall 2025 the team decided to build an AI assistant — competitors were shipping similar features, and there was already a working internal prototype from the PM: a chat that interacted with the chart.
TakeProfit is a modular trading interface. In fall 2025 the team decided to build an AI assistant — competitors were shipping similar features, and there was already a working internal prototype from the PM: a chat that interacted with the chart.
That prototype became the starting point for a full-scale feature.
That prototype became the starting point for a full-scale feature.
02. One early decision shaped how the AI would live inside the platform
02. One early decision shaped how the AI would live inside the platform
The central question: should the AI live as a platform widget or as a separate floating button? This one decision shaped everything that followed.
The central question: should the AI live as a platform widget or as a separate floating button? This one decision shaped everything that followed.
Looking back: I argued in UX terms. I should have framed it in development risk and cost — the language that mattered in that room.
Looking back: I argued in UX terms. I should have framed it in development risk and cost — the language that mattered in that room.
MY PROPOSAL
MY PROPOSAL
Widget
Widget
Native to the platform architecture, consistent with existing patterns, easier to integrate.
Native to the platform architecture, consistent with existing patterns, easier to integrate.
CHOSEN — CEO
CHOSEN — CEO
Floating button
Floating button
Positioned as a distinct product. More prominent, but harder to develop — later a source of friction.
Positioned as a distinct product. More prominent, but harder to develop — later a source of friction.
03. We built v1 around giving AI direct context from the chart
03. We built v1 around giving AI direct context from the chart
The core of v1 was a context model: the user adds a chart state — ticker, indicators, drawings — directly to a message. The AI works from that snapshot.
The core of v1 was a context model: the user adds a chart state — ticker, indicators, drawings — directly to a message. The AI works from that snapshot.
I mapped all the edge cases: what happens when a widget is removed mid-session, how the AI signals what actions it is taking, what happens when context is stale.
I mapped all the edge cases: what happens when a widget is removed mid-session, how the AI signals what actions it is taking, what happens when context is stale.
I also designed forward-looking pieces — model selection, indicator code generation via AI — knowing they wouldn’t ship in v1 but wanting the architecture to accommodate them.
I also designed forward-looking pieces — model selection, indicator code generation via AI — knowing they wouldn’t ship in v1 but wanting the architecture to accommodate them.
Shipped in v1: Chart context only. Model selection and code gen deferred.
Shipped in v1: Chart context only. Model selection and code gen deferred.
04. Real integration exposed problems we couldn’t see in design
04. Real integration exposed problems we couldn’t see in design
The problems weren’t visible in design — they emerged when real systems met real constraints.
The problems weren’t visible in design — they emerged when real systems met real constraints.
01
01
Chart architecture — elements could be added but not removed. Actions the AI took couldn’t be undone.
Chart architecture — elements could be added but not removed. Actions the AI took couldn’t be undone.
02
02
Performance — visual effects (blur, translucency) were too expensive for the runtime environment.
Performance — visual effects (blur, translucency) were too expensive for the runtime environment.
03
03
LLM output — tables, emoji, inconsistent structure broke the UI in unpredictable ways.
LLM output — tables, emoji, inconsistent structure broke the UI in unpredictable ways.
04
04
Unstable actions — the AI didn’t reliably execute chart actions, making the core promise unreliable.
Unstable actions — the AI didn’t reliably execute chart actions, making the core promise unreliable.
“By March it was clear: the problem wasn’t the interface. The AI wasn’t delivering on the product’s core promise.”
“By March it was clear: the problem wasn’t the interface. The AI wasn’t delivering on the product’s core promise.”
v1 launch was cancelled.
v1 launch was cancelled.
05. We rebuilt v2 around specialised tools instead of a single prompt
05. We rebuilt v2 around specialised tools instead of a single prompt
Instead of a single prompt, v2 introduced a system of specialised tools. The AI selects the right tool for each task, and the user can see the reasoning behind each response.
Instead of a single prompt, v2 introduced a system of specialised tools. The AI selects the right tool for each task, and the user can see the reasoning behind each response.
What I added as a designer:
What I added as a designer:
→
→
Clarifying questions — the AI asks before acting, reducing vague or incomplete responses.
Clarifying questions — the AI asks before acting, reducing vague or incomplete responses. — the AI asks before acting, reducing vague or incomplete responses.
→
→
Context signals — explicit UI cues when the AI lacks the data it needs.
Context signals — explicit UI cues when the AI lacks the data it needs. — explicit UI cues when the AI lacks the data it needs.
→
→
Model selection — finally shipping the forward-looking piece from v1 design.
Model selection — finally shipping the forward-looking piece from v1 design. — finally shipping the forward-looking piece from v1 design.
Released: April 2026
Released: April 2026
06. The rebuilt AI now handles four core trading tasks
06. The rebuilt AI now handles four core trading tasks
4
4
core capabilities
core capabilities
~10
~10
months, v1 to ship
months, v1 to ship
Apr ‘26
Apr ‘26
release date
release date
→
→
Chart analysis
Technical analysis, pattern detection, support/resistance levels, indicator interpretation, and market structure assessment
→
→
Risk management assistance
Risk calculations, position sizing, scenario analysis, and trade evaluation
→
→
Platform Guidance
TakeProfit features, chart widgets, alerts, drawing tools, and workspace setup
→
→
Indicator code generation
Writing custom indicators and strategies in TakeProfit’s Indie language, including migrations from Pine Script
What I took from designing an AI product
What I took from designing an AI product
Three things that shaped how I work now.
Three things that shaped how I work now.
01. Ship to learn, not to perfect
01. Ship to learn, not to perfect
The key insight — LLM quality wasn’t good enough — could have been discovered months earlier with a leaner test. We waited too long for a complete design.
The key insight — LLM quality wasn’t good enough — could have been discovered months earlier with a leaner test. We waited too long for a complete design.
02. Argue in the language of the room
02. Argue in the language of the room
In the widget vs. button debate I made a UX argument. The decision-maker needed to hear risk and development cost. Same logic, different frame.
In the widget vs. button debate I made a UX argument. The decision-maker needed to hear risk and development cost. Same logic, different frame.
03. Design from a live prototype
03. Design from a live prototype
v2’s approach — live prototype first, then design, then integration — was significantly more effective than v1’s linear spec-driven process.
Ann Latu · Product Designer
anna.latukhova@gmail.com
Ann Latu · Product Designer
anna.latukhova@gmail.com