ML Engineer (ONSITE IN SF)

Dear applicants, please keep in mind that applications without provided salary expectations and active LN profile will not be considered. Hope for your understanding.Location: San Francisco, CA (In-person)Employment Type: Full-TimeEquity: 0.5% – 1%Visa: Not availableExperience: 1+ years (exceptional new grads welcome)We are hiring ML Engineers to implement research ideas reliably and operate full training pipelines end-to-end. This is not a research-only role. This is research-engineering at scale. A seed-stage research-driven ML company focused on mechanistic understanding of model architectures and optimizers.The team studies:Optimizer–architecture co-designOrthogonalized optimizers and manifold-based trainingSparse attention mechanicsData-efficient reasoning modelsLearning dynamics in data-sparse regimesThe environment blends academic rigor with industrial compute and speed. The team is deliberately long-term oriented and avoids premature commercialization pressure.You will:Translate research papers into working PyTorch/JAX implementationsRun distributed transformer trainingDebug divergence and instabilityOptimize throughputBuild full pipelines (data → training → evaluation)Reason about learning dynamics and architecture tradeoffsThe bar is slope and research intuition, not years.What You’ll OwnReliable implementation of novel architecturesDistributed transformer training at scaleTraining stability and performance debuggingEvaluation frameworksOptimization reasoning alongside researchersMust-Have RequirementsStrong PyTorch or JAX proficiencyHands-on transformer training experienceExperience with distributed training setupsDebugging divergence and instabilityAbility to read and implement research papersResearch intuition around optimization and learning dynamicsHigh growth slopeNice to HaveMegatron-LM, DeepSpeed, xformersEnd-to-end pipeline ownershipResearch-engineering team experienceMathematical depth (optimization, information theory, etc.)Competitive programming / theory-heavy background