Novyte Materials - Founding AI Engineer Application

Time Required: 20–30 minutes. This is a founding role, and detailed responses are important to us.


Novyte Materials is building a frontier AI engine for materials discovery, a system that learns the structure, physics and governing rules of matter, and uses that intelligence to propose new materials for real-world synthesis.

This is not a standard ML role.

As a Founding AI Engineer, you will design generative, predictive and physics-aware architectures that operate in one of the hardest spaces in scientific ML. Your work will define the intelligence layer of Novyte’s platform.

You will build systems where:

  • Datasets are small, noisy and multimodal

  • Physics constraints are as important as the loss function

  • Exploration is as critical as prediction

  • Uncertainty is the signal, not the noise

  • Models interact with simulations and real labs

  • Rpresentation learning determines what the engine can even see

  • Every decision affects downstream experiments

We care much more about how you think than your title or background.

Please answer the questions below with as much detail as you’d like.

Read more about the job here : https://docs.google.com/document/d/1yPo7wjWyShWofHK_u9BVgMYDGDebVgF8nJvnuaAoMLg/edit?usp=sharing


Compensation  : Cash + ESOPs

Email *
1. Full Name *
2. Email Address
3. Phone Number *
4. Link to Resume/CV (GitHub, Drive, Portfolio, etc.) *
5. Link to GitHub, Papers or Technical Work (if available)
6. Which areas of frontier ML excite you most and why? (Graph learning, geometric ML, generative models, physics-aware networks, uncertainty modeling, active learning, scientific reasoning, RL for structured action spaces, etc.)
Select Relevant Areas
Graph learning
Geometric ML
Generative models (diffusion, flows)
Physics-aware networks
Uncertainty modeling
Active learning
Scientific reasoning
RL for structured action spaces
Other
If other then please elaborate
Explain your fascination and the 'why' behind your choices. *
7. Describe the hardest ML system you’ve ever designed end-to-end. Focus on: architecture selection, representation choices, dealing with noisy/non-standard data, scaling, trade-offs, what broke and how you fixed it, and what you would do differently now. *
8. Imagine you need to predict a material property, but the experimental data is sparse, inconsistent and collected under varying conditions. Walk us through your modeling strategy (preprocessing, expected biases, priors, architectures, incorporating physics constraints, and evaluating uncertainty). *
9. You train a GNN on materials; validation loss flatlines. Outline your debugging path (initial checks, inspecting representations, testing structural defects, isolating bad labels, checking symmetry, evaluating architectural expressivity). *
10. You must design an active learning loop that decides which new candidate the lab should synthesize. Explain how you would balance exploration vs. exploitation, uncertainty quantification, cost, diversity vs. refinement, and risk of model biases. (Bonus points for a concrete acquisition function) *
11. How would you incorporate physical constraints or scientific priors (e.g., symmetry, invariances, conservation laws, structural validity) directly into a neural architecture? Describe encoding these into models, architectures, or loss functions. *
12. Describe a time you built something important with no clear instructions, no clean data and no guarantee of success. What was the environment like? How did you make decisions? How did you evaluate whether to continue? What did you learn? *
13. Rate your comfort level with building production-grade ML pipelines in scientific or ambiguous domains. (Scale 1-10)
Minimal Comfort
Expert/Founding Level
Clear selection
Briefly justify your rating.
14. Which of the following scientific machine learning concepts do you have practical experience implementing? *
Required
If other,  Please Elaborate
15. Describe a specific challenge you anticipate in transitioning an AI model from a simulation environment to guiding real-world lab synthesis in materials science. *
16. When evaluating novel materials proposed by a generative model, rank the following criteria in order of importance for a successful first-pass screening (1 being most important): *
1 (Most Important)
2
3
4 (Least Important)
Predicted stability/synthesizability score
Uncertainty/novelty of the prediction
Cost-to-synthesize estimate
Diversity from known materials
A copy of your responses will be emailed to .
Submit
Clear form
Never submit passwords through Google Forms.
This content is neither created nor endorsed by Google. - Terms of Service - Privacy Policy

Does this form look suspicious? Report

Google Forms