Research Engineer - Machine Learning and Systems

Location: New York City Office HQ

Employment Type: Full time

Department: Research

Overview

We are hiring a principal level Research Engineer with deep strength in machine learning or 3D graphics, software engineering, and systems design. You will bridge frontier research with production systems and ship advanced models used in real products. The work spans exploration, rapid prototyping, rigorous experimentation, and dependable production deployment. Expect to push the limits of spatial intelligence and controllable graphics while keeping systems robust, scalable, and cost efficient.

Role

You will partner with research, engineering, and product to design, build, and operate large models and high performance systems. You will set technical standards, mentor others, and raise the bar for research quality, code quality, and reproducibility.

Key responsibilities

1. Research, design, and implement models and systems across vision, generative modeling, simulation, rendering, and 3D perception
2. Build data, training, evaluation, and deployment pipelines with strong observability and reproducibility
3. Translate research insights into reliable production services that meet product and latency requirements
4. Contribute hands on across prototyping, optimization, integration, and scaling
5. Survey new methods and run grounded evaluations to identify what to adopt and when
6. Share expertise through design reviews, mentoring, and documentation

Minimum qualifications

1. PhD in Computer Science, Machine Learning, Computer Graphics, Computer Vision, or related field, or equivalent research track record
2. Seven or more years of experience in applied ML or research engineering including significant time in fast paced or startup settings
3. Strong publication record in top venues such as NeurIPS, ICLR, ICML, CVPR, ECCV, ICCV, SIGGRAPH, or TOG with multiple first author papers or equivalent impactful artifacts
4. Proven experience training and serving large models at scale including multi GPU or multi node training, distributed data loading, mixed precision, and memory optimization
5. Fluency in Python and C++ and experience writing efficient CUDA or Triton kernels
6. Expertise with PyTorch or JAX and modern tooling for experiment tracking, evaluation, and deployment
7. Demonstrated ability to take ideas from paper to production with measurable impact on users or business outcomes
8. Strong systems skills including profiling, performance tuning, reliability engineering, and cost awareness
9. Excellent communication with the ability to work across research and product teams

Preferred qualifications

1. Contributions that are widely used in the community such as open source libraries, datasets, or benchmarks with visible adoption
2. Experience in neural rendering, differentiable rendering, 3D reconstruction, volumetric video, SLAM, geometric deep learning, or simulation
3. Experience operating large training jobs on Kubernetes, Slurm, or Ray across public cloud environments
4. Experience with evaluation and safety for generative or interactive models including red teaming and guardrail design
5. Track record of mentoring teams and setting research and engineering best practices
6. Patents or awards that recognize technical contributions

Nice to have

1. Shipped interactive graphics or 3D systems with strict real time constraints
2. Experience building custom compilers or graph level optimizations such as CUDA graphs, XLA, or graph capture
3. Prior leadership in cross functional initiatives spanning data, infra, and product

How to apply

Please include a CV, links to publications, code, and a brief summary of two projects that best represent your impact. Include details on model scale, data scale, latency or throughput targets, and the concrete results you achieved.

On site in New York City required. Relocation support available.

• Recruiter screen 20 minutes
  Goal: motivation, location in NYC, work authorization, compensation range, start date.
  Quick bar: publication record or equivalent artifacts, experience training or serving large models, Python and C plus plus comfort.
• Hiring manager deep dive 45 minutes
  Goal: end to end ownership.
  Prompts: walk through one system or model you built from idea to prod. Detail data scale, model scale, infra, metrics, cost, failures, and what you would change now.
  Signals: clarity, decision quality, tradeoffs, actual impact.
• Portfolio and publications review 45 minutes
  Prework: panel reads two of the candidate’s papers or equivalent artifacts.
  In session: candidate defends novelty, ablations, limits, reproducibility, and lessons.
  Signals: research rigor, originality, evaluation hygiene, ability to ship.
• Live coding 70 minutes
  Format: one practical problem, candidate chooses Python or C plus plus.
  Options to pick from
  a. Build a minimal training loop with correct seeding, mixed precision, gradient clipping, and a tiny metric dashboard.
  b. Write a performant kernel style function for a fused op or a small 3D primitive, then profile and optimize memory and throughput.
  Scoring: correctness first, then profiling, then maintainability and tests.
• Systems design and scalability 60 minutes
  Prompt: design a training and evaluation pipeline for a spatial intelligence model used in production with daily model refreshes. Include data curation, feature stores, versioning, distributed training, eval slices, rollout, guardrails, on call plan, and cost model.
  Signals: architecture under real constraints, reliability, cost awareness, observability.
• Research talk 30 minutes plus 15 minutes Q and A
  Candidate presents recent work of their own.
  Signals: depth, taste, problem framing, ability to teach complex ideas.
• Graphics or 3D focus round 45 minutes
  Candidate chooses one
  a. Neural rendering and differentiable rendering fundamentals
  b. 3D reconstruction and spatial perception
  c. GPU performance engineering and memory models
  Signals: genuine expertise in at least one of these areas.
• Product and collaboration case 40 minutes
  Prompt: a PM and infra engineer describe a product goal with strict latency and cost targets. Candidate proposes a research and delivery plan with milestones, offline to online metrics, success criteria, and de risk experiments.
  Signals: translation from research to product, prioritization, stakeholder management.
• Leadership and values 30 minutes
  Topics: raising the bar, mentoring, code and research standards, handling setbacks, authorship ethics, open source posture.
  Signals: team builder, owner mindset, integrity.
• Writing assessment async 45 to 60 minutes
  Prompt: write a one page plan to improve quality or latency for a spatial model. Include hypothesis, experiment design, metrics, risks, and a rollback plan.
  Signals: crisp written communication, experimental rigor.
• Reference checks three calls
  Who: former manager, senior peer, cross functional partner.
  Focus: independence, technical bar, collaboration, mentoring, delivery under ambiguity, reliability in production.
• Decision and calibration
  A single rubric with four outcomes for each dimension: strong hire, hire, no hire, strong no hire.
  Dimensions
  a. Research excellence and publications or equivalent artifacts
  b. Systems and performance engineering
  c. Training and serving large models at scale
  d. Graphics or 3D depth
  e. Product impact and judgment
  f. Communication and leadership
  Hire requires strong hire on at least two technical dimensions and no lower than hire on the rest.