Software Engineer2017—2018
200K+
daily users
Dabbl
Led the frontend for ShopRite's Downtime Dollars campaign and shipped the client work that earned a follow-on Kroger contract.
San Francisco · Product + AI/ML · 2026
The route was never linear.The logic was.
Justin Chang builds systems for the moments when the problem is still unclear.
9
years building products
SWE → ML
deliberate evolution at Meta
15%
model-action lift
200K+
daily users served
01The route
The long way to software.
The unconventional decisions started early. Each turn looked separate at the time; together, they became a way of thinking.
Look for the overlooked path. Then do the work.
Field note 01Age 16CHSPE · Community college
A calculated exit
I found a nontraditional route through California's education system, studied the probabilities, and committed. The bet led to admission at Berkeley, UCLA, and UC San Diego.
2013UCLA · Philosophy
Logic, in human form
Philosophy felt like the most mathematical corner of the humanities—a way to sharpen logic, reading, and writing while preparing for medicine.
2014Columbia · Pre-med
The serious plan
Medicine was not a casual idea. I entered Columbia's post-baccalaureate program to complete the science required to pursue it fully.
2015–16SEO writer · Between paths
The honest detour
I learned what kind of life I did not want. For two unsettled years, programming kept returning—part necessity, part fascination, all possibility.
2017Software · First products
The world opened
Code turned ideas into artifacts. The appeal was bigger than a job: with the skill, the number of futures I could build seemed to expand.
02Learning the craft
The gap became the curriculum.
San Francisco revealed a higher standard. The response was simple: get close to stronger engineers, ship for real users, and close the distance.
Software Engineer2018—2019
+30%
subcontractor MAU
BuildingConnected
Moved into San Francisco's engineering culture, saw the gap clearly, and doubled down until measurable product impact followed.
“The first lesson was how much I didn’t know. The second was that I could close the gap.”
03Scale changed the stakes
Confidence turned into a plan.
Three teams. A promotion for exceptional performance. A deliberate move from software engineering into machine learning before the next bet began.
Menlo Park · 2019—2023
“I wanted to know whether I was good among great engineers.”
The answer created enough confidence to plan the next four years with intention.
2019—2020
Recruiting Products
10+ hours reclaimed weekly
2020—2022
Ads Metrics
0% → 80% audit coverage
2022—2023
ML Integrity
+15% model action rate
Promoted
Consecutive high-performance cycles
0 → 80%
Ads metric audit coverage
+15%
Action rate for severe-harm models
04The bet
Then I bet on my range.
Two years building independently turned product instincts, software, ML, writing, sales, and operations into one operating system. The combination was the advantage.
person · explaining
dog · inspecting
Human + dog action modelLive inference
A.01 · Computer vision · Flagship build
Action intelligence
Custom human and dog action models, optional identity association, and a production system that reached two beta users. VLM-assisted annotation cut the expensive part—the dataset—by one to two orders of magnitude.
10–100×
lower annotation cost
A.02Applied LLMs · GPT-3.5 era
4,096Context ceiling
ProductImagesSeasonHistory
Context beyond the window
A social and long-form content platform that understood products, images, seasons, and offers—then produced coherent multi-page technical writing inside a 4,096-token world.
A.03Client systems · Revenue
Product
Leads
Systems
The product behind the business
The first product created leads. Those customers revealed harder, more valuable problems, and custom product work became the company's commercial engine.
02 yrsbuilding independently
Lab notes
Small experiments. Fast loops. Ideas made tangible before they had permission to become products.
L.01
Action Labeler
VLM autodistillation
L.02
Glimpse
Vision Pro timeline
L.03
Advenshare
Spatial simulation
05The frontier
Turn noise into signal.
At OpenAI, the setting changed. The pattern did not: find the ambiguous problem, build the system, and make its value visible.
External intelligence · Integrity
Evidence becomes a product.
A multi-source intelligence platform for the people who need to understand what is happening, why it matters, and what changed.
03 mo
to breakthrough
Org-wide
initiative
Member of Technical Staff · OpenAI
“Find the ambiguous problem. Build the system. Make its value visible.”
06The method
The pattern underneath the work.
Different industries. Different technologies. The same operating principles keep resurfacing.
Make the abstract concrete.
A working product is often the clearest argument.
Own the whole system.
Interface, model, data, infrastructure, operations, and adoption are connected.
Let range compound.
Writing, philosophy, product judgment, software, and ML improve one another.
Learn at the scale of the ambition.
Acquire the missing capability before the next bet requires it.
Measure impact through use.
Adoption and changed behavior matter more than novelty alone.
07Away from the screen
Still building. Just breathing differently.
Running, tennis, the beach, Muir Woods, and a dog who keeps becoming the inspiration—or product manager—for the next experiment.
The next field test: camping. Then backpacking.
Off-hours operating system
TennisRunningThe beachMuir WoodsDog projects
The boundary between life and building is porous. Practical problems become experiments; experiments sometimes become products.
08 · What comes next
Build enough of the answer to make the future visible.
The work I want next is where the answer is not obvious yet—and where building it is the fastest way to find out.
Product leadership · Technical depth · Ambitious teams
Justin Chang · San FranciscoBack to beginning ↑