ML Research Engineer

About the Role

We are looking for an engineer to drive the models behind our design generation, starting with Vyb. You will construct data collection pipelines, train models, and design evals that measure what designers actually care about. Your focus will be on turning ambiguous quality problems (e.g. this design does not feel premium) into measurable objectives and shipping models that move them.

Our Core Engineering Philosophy: We highly value simple solutions to complex problems. We want engineers who reason from first principles, run the smallest experiment that answers the question, and prefer clean fixes. If you take pride in rigorous experimentation, reproducible pipelines, and stripping away complexity rather than adding it, you will fit right in.

What You Will Do

• Model Training: Design, run, and iterate on training for our generation models. Make pragmatic calls on model architectures and training strategies, and know when a problem calls for more data, better data, or a different objective.

• Build Synthetic & Curated Data Pipelines: Build the systems that generate, render, validate, and filter training data at scale. You will treat data quality as the primary lever on model quality.

• Design Evals: Build the evaluation frameworks that turn subjective design quality into metrics we trust, combining automated checks, model-based judges, and human preference signals.

• Make Models Production-Ready: Optimize inference latency and cost and minimize training / serving skew. Keep experiments reproducible, versioned, and easy to compare. Write clean, tested training and eval code that teammates can read, rerun, and debug without archaeology.

What We Look For

• 3–5 Years Training Models That Shipped: You have trained, fine-tuned, and evaluated models that made it into real products, ideally at a scaling startup or a strong applied ML team.

• Obsession with Simplicity & Measurement: A proven track record of cutting through ambiguous problems with simple baselines and clear metrics. You don't trust a result you can't reproduce, and you don't trust a model you can't measure.

• Strong Modern ML Fundamentals: Deep hands-on experience with PyTorch and the transformer stack, including fine-tuning (SFT and RLHF), distributed training, and debugging loss curves, data bugs, and silent failure modes.

• Data-Centric Instincts: Experience building large-scale data generation and filtering pipelines. You know that most model problems are data problems, and you're comfortable writing the deterministic validation and filtering code that proves it.

• Multimodal & Structured Output Experience: Familiarity with vision-language models, embeddings, and generating structured outputs where validity and correctness matter as much as fluency.

• Work Authorization: Applicants must be currently authorized to work in the United States on a full-time basis. We are unable to sponsor or take over sponsorship of employment visas (e.g., H-1B) at this time.