Founding Engineer - ML Research

About Us

Deccan AI is a model training and eval startup headquartered in the Bay Area. We are founded by IIT Bombay, IIM Ahmedabad and ex-Google alumni, and work with some of the top AI frontier labs in the world, e.g. Google Deepmind, Snowflake, and many more. We are backed by Prosus Ventures, and our India delivery office is in Hyderabad. Most of our roles are on-site i.e. Bay Area / Hyderabad.
About the Role

We’re looking for a Founding ML Research Engineer to help build and scale our AI research backbone from the ground up. This role sits at the intersection of applied machine learning and systems engineering — ideal for someone who thrives in unstructured environments, loves building fast, and obsesses over model performance and data quality.

You’ll work directly with the founding team to prototype, experiment, and ship research-driven features that push the boundaries of what our AI systems can do.

• Design, train, and evaluate ML models (LLMs, diffusion models, or domain-specific architectures).

• Develop scalable experimentation pipelines for data, model, and evaluation workflows.

• Work closely with the data and infra teams to optimize training throughput and quality.

• Contribute to open research, internal benchmarks, and emerging techniques in multimodal or generative AI.

• Rapidly prototype and productionize research insights into usable tools and models.

• Define foundational technical culture — set standards for research rigor, documentation, and reproducibility.

• 3+ years of experience in ML research, applied ML, or ML systems engineering.

• Deep familiarity with PyTorch / JAX / TensorFlow, and model architectures (Transformers, Diffusion, or RLHF).

• Strong foundations in data processing, distributed training, and evaluation metrics.

• Curiosity for emerging ML paradigms (multimodality, self-learning, synthetic data, agentic systems, etc.).

• Ability to move from research papers → working prototypes → production-ready code.

• Passion for building from zero to one in a high-velocity startup environment.