← hifrieda.com
exchange 01
you · 11:02How do you use AI in your design process?
frieda ✦I don't use it to make pictures faster. I use it to hold the flow, the edge cases and the states on one canvas — here's the tool I built for that.
spark.jsxlive
selected work

WHOOP

lead design · WHOOP AI

Codentist

founding designer

Eaglenos Health

CGM platform

Red Cross simulation

with CMU's SDM lab

shanghai ↔ boston

exchange 02
you · 11:04Show me the dental work
frieda ✦Codentist. I was the founding designer on an AI scribe for dentists — v1 shipped in six weeks.
codentist
frieda ✦ connecting a few dots…

A portfolio that reflects my identity shouldn't make you hunt for the answer.

drag the board around

the whole board
100%

01

learn more about
me personally

Don't navigate my portfolio. Ask it.

Most portfolios make the visitor do the retrieval. You land on a wall of projects, open a few, and try to figure out which one answers the question you actually came with.

I wanted to reverse that.

Ask what you want to know, and the portfolio brings back both the answer and the work that supports it. The conversation gives you the story. The artifacts beside it give you the evidence.

Try a question ↓

frieda ✦

I don't use it to make pictures faster. I use it to hold the flow, the edge cases and the states on one canvas — so I built the tool that does it.

why these → she builds the tools she wishes existed, in code, and uses them on real work

Four of the questions people actually ask.

02

context, gathered
in one place

Design work doesn't fit down a single column.

The first version of this portfolio was a chat window.

A thread is optimized for turn-taking: one question, one answer, then the next. But design work depends on relationships. You want the previous screen beside the next one. The prototype beside the explanation. It's how designers naturally build visual context and connect the dots.

So I stopped treating the conversation as the interface. On the canvas, every exchange becomes an object in a shared space. Answers, screens, prototypes, and the questions that surfaced them stay visible together. You can pan across the thinking instead of reconstructing it from a transcript.

2 of 4 asked

v1 — a chat window

ffriedaonline
Ask something…
One lane. Each piece at bubble width, only the last one in view.

now — the canvas

exchange 01
youHow do you use AI in your process?
frieda ✦Flow, edge cases and states — on one canvas.
spark.jsxlive
exchange 02
youShow me the dental work
frieda ✦Codentist — founding designer on the AI scribe.
codentist
The same four, side by side and comparable at a glance. Hover to step onto the board.

03

every answer
gets graded

The canvas is only what you see. The agent logic is what makes it useful.

Behind the interface is a system that decides what the agent knows, what it retrieves, how it responds, and what should happen next.

Projects, artifacts, skills, and personal context are structured so the agent can reason over them. Prompting defines how it interprets a question, what evidence it should pull in, and how strongly it should make a claim.

From there, the flow is automatic: understand the intent, retrieve the most relevant context, compose the response, attach supporting work, and place the result back onto the canvas.

The most important layer is evaluation. I built an eval module into the backend: metrics that score the quality of each response, and the signals they surface are what I use to adjust the agent's responses and its prompt strategy.

ConsoleFrieda Chen

Evals run 14

A sample run, graded on the same rubric as the live agent.

agent Sonnet 5effort mediumjudge Opus 56 cases
8.5/10 overall

    0 fabrications

      Suggested actions

      A worked example. The tabs mirror the live console; the companies, counts and scores are illustrative, and the rubric, judge and kinds of action are the real ones.