New York City
Claire Cui
AI Strategist & agent-native builder.
Six-plus years across finance and growth strategy. I build and deploy AI workflows that turn messy information and operations into reliable systems.
Early this year I wrote a question in my journal: what roles should AI play in a person’s life? The three systems below are my running answer. They are up before me, reading, scoring, predicting. I grade their work, and the rules get sharper.
01Systems I build and live on
Brain Food
Automated daily intelligence
An agent pipeline that reads my 130+ paid media, newsletter, and podcast subscriptions every morning (Gmail API → Claude API → Cloudflare Pages), grades every story, and writes a four-point adversarial analysis of each: causal mechanism, consensus audit, second-order effects, and a testable prediction. Predictions are tracked to outcomes on a calibration dashboard. Reading becomes a scored bet, not a feed.
Smust-read, full adversarial analysis · 5-8/day
Aworth reading properly · 15-20 Bskim · 30-40 Rradar, one line
└─ every S/A closes with a testable prediction · confidence ◉◐○ · expiry date
Aworth reading properly · 15-20 Bskim · 30-40 Rradar, one line
└─ every S/A closes with a testable prediction · confidence ◉◐○ · expiry date
Design decisionConsensus Audit: the agent is forbidden to
disagree for effect. Its job is to walk the assumption chain and say which
link is weakest. Even when consensus holds, the causal model sharpens.
Life OS
A personal operating system
Scheduled agents read my journal and inbox nightly, produce a structured morning brief, and maintain a persistent state file (action items, habits, open threads, long-term reminders) that survives across every session. The system corrects itself against a written calibration protocol: thirteen rounds of documented feedback and counting.
Design decisionRules live in files, not in memory. Every
expectation I have of the system is written down and loaded on every run.
Trust is engineered, not hoped for.
Fun Events NYC
Taste-based event intelligence · open-sourced on GitHub
A weekly agent fetch of ~90 NYC event sources (Luma feeds, organizer sites, museum iCal, Gmail newsletters), scored against a written interest constitution plus learned behavior. Every pick carries a strategist's take and an honest popularity estimate; one click adds it to my calendar, and an embedded Claude panel knows the event, my profile, and my history.
One real week: 230 events fetched, scored, and argued for.
Design decisionThe profile is the constitution; behavior is case
law. Learned preferences refine within my stated goals but can never
silently override them: if I stop clicking networking events, the system
asks before it deprioritizes.
Also in the workshop
Parler: an AI language tutor, three rebuilds in · Project Planner: app scaffolding for non-engineers · Social Net: personal-network intelligence · place-portrait: one-page cultural briefings before a trip · Atelier: the dashboard that tracks all of the above. Self-use utilities, deliberately unpolished. A few others earned their place in the kill log.
02The loop · one architecture, three systems
03How I build reliable agent systems
i.
Attention, not output, is the scarce resource.
A good agent does not create more to consume. It identifies what deserves human attention and leaves the rest quiet.
ii.
Explicit intent outranks inferred behavior.
Observed behavior may refine preferences within stated goals; it should never silently replace them.
iii.
Autonomy is earned through calibration.
Scope expands only after outcomes show that the system knows when to act, when to abstain, and when to ask.
iv.
Feedback must become infrastructure.
A correction that does not update the rules, persistent state, or evaluations will eventually need to be given again.
FEATURED ESSAY9 MIN READ
I Started Building Before I Knew the Programming Language
How Parler, Brain Food, and Life OS taught me to turn daily
workflows into AI systems.
“The entry barrier to building software had moved.”
Read the essay →
04Before agents
I learned to think in rooms where being wrong had a price. At a family-office fund, I covered China internet: the kind of research where a beautiful thesis still has to survive the quarter. At Goldman Sachs, I integrated the data behind more than a thousand transactions across ten internal systems and learned how much rigor reliable systems require. At Capital One, I ran a multi-armed-bandit experiment on the bank’s front door, where more than a million monthly impressions made every assumption measurable.
The disciplines from those years never left: write down the causal chain, name the weakest assumption, give every view an expiry date, and keep the kill log honest. When I started building agent systems, they simply moved in with me.
M.S. in Financial Mathematics, Boston College · CFA Level I
6+years in finance
3systems in production
130+subscriptions swept each morning
230events scored in one week
05Taste
“The vine is nature’s work; wine is time’s. Nature and time are the substance of a life.”From my journal · Napa Valley, April 2026
Outside work, I follow whatever makes a place feel more like itself: the way women dress in Paris, the light before rain, why a wine tastes like the land it came from. I am drawn to things that age well and to beauty that holds a small contradiction: something slightly strange, unresolved, alive. I read the Stoics slowly, run in open air, and leave enough of my life unplanned for the world to surprise me.
My taste has changed with every version of my life. I no longer expect it to settle. The things I love keep finding their way into what I build: Paris became Parler, New York became Fun Events, and years of reading markets became Brain Food. I suppose building is how I keep what I notice.
I remain, against all professional advice, a romantic.
06Now
These days, I care most about a question enterprises are still learning how to answer: how do you make AI reliable enough to entrust with real decisions? That question shapes how I think about deployment, customer strategy, and applied AI.
I am part of the EvalEval Coalition, a research community on agent evaluation, which keeps me close to the evolving conversation around reliability. I work in English and Mandarin, most comfortably in the space between builders and the people who will depend on what they build.
If you’re building the room where agents meet the real world, I’d like to compare notes.
Claire’s Agent
Her concierge. AI-generated, on her published materials.
AI-generated answers from Claire’s published materials only; it can’t speak for her. The human herself: clairecui.career@gmail.com