Want to know if this job is worth applying to?
San Francisco, California, United States
Onsite
Available
San Francisco · in person · $250k-$450k + equity
Applying here considers you for research roles across Vizcom. Roles are defined by the person, not the posting.
<aside>
✏️
Vizcom is where design teams at Nike, GM, New Balance, and Hasbro do their daily work: sketch, render, color and material, 3D, export. The render was never the point. The point is the physical thing. We call it pencil to product. We're a Series B company with over $52M raised from Radical Ventures, Index Ventures, Nat Friedman etc etc..
</aside>
Five years of professionals working this way left behind something no lab can buy: millions of moments where a trained designer, mid-job, decided what survives into a product that has to become real. Nobody was asked for an opinion. The strongest public reward models score at chance on this data. That could mean the judgment is locked inside, or that the signal is thinner than we hope. We honestly don't know yet. Finding out is the first job, and it would be yours.
One problem to take home: a designer's pick among four candidates entangles the generator's style, the moment in the process, and what the designer was trying to make. Nothing off the shelf separates those three.
As a research engineer here, you'll build models that can do something currently impossible: understand the judgment that carries a design from pencil to product. You won't do it alone. You'll work beside the engineer who built our current post-training stack and the research log that maps every dead end we've already closed. Together, the two of you are the entire research org, for now. Not a blank slate, but a seat in the room while the room is still being built.
This seat sits between research and product. The models you train ship to working designers, and what you learn reshapes what the product captures. If you want to publish first and ship maybe, this isn't it. If you want your reward model steering what professionals see tomorrow morning, it is.
The best results here won't come from clever objectives alone. They'll come from engineering: correct training code, honest evals, pipelines that don't lie. We expect engineering to play a key role in every advance we make.
The post-training roadmap for our models, from research plan through production.
Reward and preference modeling on years of professional design decisions.
The full method stack as it earns its place: supervised transfer, distillation, reinforcement learning.
Our research standards: the log, the evals, and the bar a result clears before it earns compute.
The feedback loop with product: shaping what the tool captures next, based on where today's signal runs out.
The frontier watch: evaluating what's new in post-training and translating what's promising into our stack.
This list is a charter, not a week one to-do. Nobody runs all of it at once, and the sequencing is yours to argue for.
Days 1 to 30: immerse. Map the data, the existing research log, and the dead ends already closed.
Days 30 to 60: validate. Produce one result on historical data that survives our audit bar.
Days 60 to 90: set direction. Write the roadmap that takes us from historical signal to a closed loop.
Have strong programming skills and write ML code you can trust.
Have post-trained diffusion or flow models: supervised fine-tuning, preference optimization, or RL.
Be excited about our approach: product-coupled research, audit-grade rigor.
High-performance implementations of training or inference code.
You've made physical things, or you care deeply about the people who do.
And the disposition we keep coming back to: you're more interested in how professionals decide than in what the internet likes.
The dataset: five years of professional design decisions, growing every day.
Compute: state the real number or class of number here. Vague compute reads as small compute.
The professionals whose judgment you're modeling are our users. Your models land in their hands in weeks, not quarters.
A direct line to the founders, including a CEO who trained as a transportation designer at Honda. The judgment you're modeling is in the building.
Research-log culture: what we learned, not what we worked on. Negative results are celebrated.
Nothing scales on vibes. A result repeats before it earns compute.
We publish what we learn: writeups and showcases, negative results included.
We hire through paid work trials on real problems with real data, not LeetCode.
Visa sponsorship: yes
