Licia Li, Lead product designer

(about)

Agentic valuations (01) Simplifying complex payouts (02) Trust in AI extraction (03)

Lead product designer at CartaSan Francisco

I lead design for Carta's valuation product, which we built from 0 to 1 and rebuilt to be AI-native in Claude using MCP and skills. I'm defining the agent ↔ human interaction model while the technology is still taking shape.

Next Simplifying complex payouts (02)

Simplifying the complex: I turned private equity payouts into a map that shows who gets paid how much, and why.

Next Trust in AI extraction (03)

Trust in AI extraction: I made every AI-extracted number in portfolio financials traceable to its source, who edited it, and how it was calculated.

Next Agentic valuations (01)

I'm Licia Li. I started drawing when I was five (still do). I trained as a landscape architect, then turned to software eight years ago. The projects I'm proudest of at Carta started without a brief: I spotted the problems, prototyped the answers, and got them built. When I'm not at a screen, I'm probably out on my road bike.

San Francisco heyliciali@gmail.com LinkedIn (↗)

Megabiblioteca (Mexico City)

Claude's chat beside the Carta valuations dashboard for Q3 2026. In chat, Carta says 24 companies are in scope and 4 are missing data, then asks how to start; Bulk run with available data is selected and Continue is ready. The dashboard shows $849.6M in holdings value. — Fig. 1The landing experience: the chat captures user intent, and the dashboard opens alongside with the context to decide.

Fig. 2Two ways to review the agent's suggestions before a bulk run. 1a Inline table: one row per company, with method weights in a popover. 1b Grouped by setup: companies with the same setup are reviewed and changed as a group.

Fig. 3Using an MCP app and tool calls to show the agent's progress visually.

Fig. 4Tinker with method weights and watch the total value update. I chose a single bar UX over constant-sum sliders because one bar makes the 100% total obvious, while separate sliders move on their own like ghosts.

Fig. 5See the high-level change in chat, then click Review details to drill in on the side panel.

Role: Design lead
Scope: Enterprise AI, Agent UX, 0 → 1
Shipped: Still shipping new features every week

Context

Investors need to value their portfolio companies, and the process has been manual and painful for years. Our 0 → 1 web app moved that work off spreadsheets in 2025. As the AI landscape changed, we rebuilt the entire enterprise-grade product in Claude for an agent to run.

My approach

Valuation is part science, part art. Agents can do both, and only ask people to review the art. The system syncs the latest inputs, drafts every company's valuation the way each customer did it last quarter, and flags what changed in a meaningful way.

I got to work with cutting-edge AI features to define the UX paradigms. My design recommendations changed over time as Anthropic launched more features: widgets, artifacts, MCP apps, in-app browsers, and the list keeps growing. While the surfaces keep changing, my design principle has stayed about the same: let agents do the complex work in one shot, then surface the changes and decisions so people can sign off faster and more confidently.

Impact

Revenue:The 0 → 1 product found product-market fit and drove ~$50M ARR
Customers:Saves days to weeks of manual work every quarter
Design:15 other Carta teams now build on the UX patterns I recommended

Fig. 1Payouts sit right on the deal structure: money moves along each line, and the side panel lists who gets paid how much.

Fig. 2From the full map, drill into a holder's path, then a single entity; the side panel always matches the map.

Fig. 3Drilling in works the same in the flow chart: follow a holder's path, then focus on a single entity.

Fig. 4The interaction details behind the map: one screen with three views, from the map to a table, and focus on one holder's path or a single entity.

Fig. 5

Fig. 6A waterfall run takes about 10 seconds: long enough to notice, too short to step away. Instead of a standard shimmer, I explored loaders that turn the wait into a small moment of delight.

Role: Design lead
Scope: Enterprise, Data viz, Motion
Shipped: Oct 2026

Context

Private equity firms have complex deal structures, which make it hard to understand who gets paid how much from an investment's return. I designed a waterfall product where people can run complex waterfalls, find the numbers quickly, and see where each payout comes from.

My approach

The backbone of the payout calculation is the deal structure, so instead of dense tables, I put the payouts right on a map of it.

Based on my user research into what context helps people find a number fast, I decided on the UX pattern: a full map you can drill into layer by layer, with just the right amount of information at each level, like Google Maps.

Motion was the fun part: it reduces cognitive load and helps people find information fast. Money visibly moves along each line, so any payout traces back to where it came from; zooming in pulls the related entities together; and hovering highlights the same entity on the map and in the table beside it.

Impact

Sales:Prospects have responded extremely well since the October launch

The source in full-page view: the BetaHex Q3 2026 spreadsheet fills the window, with Total Cost of Goods Sold for Q4 2025 boxed and its tag showing the extracted value and the current value, both $3,698,666.07. A round magnifier enlarges the boxed cell and its tag.

Five pieces on one plate: the Gross Margin tile for FY 2025, tracing Gross Profit divided by Revenue down to four quarters with Q3 2025 selected; a tile reading No source file with a manual correction in its audit log; a tile reading Entered manually; the note that annual figures are resampled from quarters; and the What looks wrong? form with Send to Carta.

Role: Design lead
Scope: Enterprise AI, Data trust
Shipped: Nov 2026

Context

Companies share their financials with their investors on Carta, where AI extracts the numbers. To trust them, investors need to trace each one back to its source.

My approach

I identified three core user needs when tracing a number back:

see a cell's edit history
verify extraction accuracy
track the formula

From there, I explored placements and user flows to match each need to the right entry point and interaction pattern.