Skip to the work
BYBorn Yesterday
A personal laboratory / Brian Chen
Selected systems / Field note 00San Francisco / 2026

Sales strategy / Revenue operations / GTM systems / Forecasting

Born yesterday.
Start with the problem, not the precedent.

Selected work

The work starts when the existing answer stops making sense.

Sales strategyRevenue operationsGTM systemsForecasting

A forecast built with knowledge from the future. A home search that ignores the photos. Markets that look comparable until you inspect the contract. Memorabilia stripped of the context that gives it value. Each project begins at the explanatory failure, then follows the bottleneck to a working system.

01 / Deepest rabbit holeData engineering / Quant research / Production systems

Case file 01

How do you know the unknowable?

Predicting the future is the most challenging and intoxicating problem I know. Survivorship is where I chose to wrestle with it.

Markets are an endless teacher because every forecast eventually meets reality. But that lesson is only useful if the experiment was honest. I had to reconstruct the investable universe, preserve what was knowable at each moment, audit vendors, engineer features, prevent leakage, design walk-forward experiments, and operate the same logic in production. The forecast is one part of a system built to make failure informative.

Why this is hardI built a point-in-time research and production system across 10M+ rows, so a result can be traced back through the universe, data, features, validation regime, and code that produced it.Forecasting / Data engineering / Experimental design / Production reliability
Explore the public framework ↗
Exhibit 01.AAnatomy of a one-person quant system
One downstream artifactModel explainability
Illustrative SHAP beeswarm using real Survivorship feature names including ftd_zscore and si_days_to_cover, with synthetic feature impacts colored from low values in blue to high values in pink
7,800+ securities10M+ point-in-time rows223 / 223 research configurations completed750 / 64 / 100 train / embargo / test dates

Public-safe evidence only. Operational data, tuned strategies, model artifacts, and live research remain private.

02 / Working filesThree different domains / One operating instinct

Three bottlenecks.
Three working answers.

The visible product is the last mile. Each case starts with the decision that fails, then shows the system required to make it trustworthy.

HHunter interface showing a property photograph, AI room analysis, and human-review evidence
Actual product interface / Listing identity withheldExhibit 02.1

Case file 02.1

HHunter

What if home search could use the evidence hidden in the photos?

Standard filters can compare price, size, and bedrooms. They discard much of the evidence that actually changes a housing decision. HHunter turns listing photos into a room-by-room review queue, uses multimodal AI to identify relevant visual signals, and attaches the source image to every judgment.

Why this is hardThe model narrows the search and shows its work. It does not pretend to choose a home.Multimodal inference / Evidence-backed review / Human-in-the-loop product
Courtline concept showing a vintage tennis racket photographed on a clay court
Illustrative product concept / Synthetic inventoryExhibit 02.2

Case file 02.2

Courtline

What makes a tennis object worth collecting?

A racket is not valuable in isolation. Athlete, moment, provenance, scarcity, and presentation create the reason to care, yet generic commerce grids flatten that context. I translated the category thesis into believable inventory, auction mechanics, buyer journeys, pricing context, and two distinct marketplace directions.

Why this is hardThe strategy had to survive contact with inventory, mechanics, and merchandising. Two working directions make the product thesis comparable rather than hypothetical.Product strategy / Marketplace architecture / Merchandising / Visual direction
View Courtline V1View Courtline V2
Actual candidate queueRead-only operator surface
Read-only candidate screener showing market filters, review counts, and cross-market opportunities
System architecture + actual candidate screenerExhibit 02.3

Case file 02.3

Cross-market candidate screener

Can you compare the same market, on the same side, before the price moves?

The opportunity exists only if contracts from different venues are truly equivalent and the data is still fresh. I shipped an operator surface early, then used the working system to expose the real bottlenecks: concurrent intake, outcome normalization, exact contract-and-side matching, and source evidence.

Why this is hardDevelopment speed made the idea testable. Bounded concurrency, cached metadata, normalized contracts, and explicit freshness made the result usable.O(4×N) → O(4) request batches / Bounded concurrency / Cached opening prices
03 / Before AIPersonal experiments / No affiliation

Before AI, the interface was the API.

The tools were different, but the operating instinct was already there: watch the condition continuously and return only when the decision changes.

Illustrative scene / active fishingReconstructed detector view
Illustrative Final Fantasy XIV scene showing a character actively fishing from a dock, with a visible rod, line, bobber, and interface
Window capture1600 × 900
ObserveSystem feedback
MatchFisher actions
WatchGP state
Template bankActual project crops
Archived bite-signal image template
01 / Bite signal
Archived successful-catch image template
02 / Catch confirmation
Archived full-GP image template
03 / Full GP
Idle→Fishing→Reeling→Verify↺
01CaptureRead the live window
02RecognizeMatch visual states
03DecideAdvance the state machine
04VerifyObserve the result
03.0Archived CV experiment

Built before general-purpose multimodal models

Teaching a machine to fish.

Can a visual interface become reliable machine state?

With no structured state to query, I used real-time window capture, OpenCV template matching, and an explicit state machine to interpret a changing Final Fantasy XIV interface. The system recognized bites, catches, failures, available abilities, and resource constraints, selected the next action, then watched the screen to verify the result.

Why this is hardEvery action had to be chosen and verified from noisy pixels alone.Real-time vision / State machines / Closed-loop automation / Failure recovery

Illustrative fishing scene paired with original 2020 image templates; detector overlay reconstructed for this case study. Archived personal computer-vision experiment. No affiliation with Square Enix. No code or operating instructions are published here.

Smaller systems / Same instinctSet the condition once
03.1Built

Too Good To Go

The itch

Certain vendors released surplus bags at fairly consistent times, but the early app did not offer the notification I needed.

What I did

I built a personal monitor for the vendors and bags I cared about. When one appeared, it sent me a Telegram message.

Exit conditionNo reopening the app just to find nothing.
03.2Built

SeatGeek

The itch

There are plenty of events I would attend if the last-minute price became cheap enough. Rechecking them all defeats the point of being spontaneous.

What I did

I could set a personal go-price, such as $40, then leave a lightweight script to watch the event and tell me when the answer became yes.

Exit conditionDecide once. Hear about it only when the price matters.
03.3Adopted

Camply

The itch

Finding the right campground meant repeatedly copying exact campsite criteria and checking for an opening.

What I did

I started building my own monitor, then serendipitously found an open-source project that already solved the problem. I used it.

Exit conditionThe objective was a campsite, not authorship.

Descriptive accounts of personal experiments. No affiliation with the named services. Camply is explicitly included as an example of adopting a useful existing solution rather than rebuilding it.

The same instinct at work

Find the actual bottleneck.
Make the system legible.
Carry it through.

I do my best work when the problem is real, the path is unclear, and the answer has to survive contact with actual users or data.

Across sales strategy, revenue operations, GTM systems, and forecasting, I move between diagnosis, system design, implementation, and evidence. Born Yesterday is where I pursue the problems I choose myself.

Email ↗LinkedIn ↗GitHub ↗