Scientific Data & Data Engine Scientist / Engineer

Scientific Data & Data Engine Scientist / Engineer

Role Overview

You will work on Scinetics’ scientific data infrastructure and Data Engine, continuously building high-quality, trainable, and verifiable scientific data for foundation models, post-training, and Scientific Agents.

We care about more than “collecting more data.” The challenge is to turn scientific information scattered across papers, databases, scientific software, agent trajectories, and experiments into training signals that genuinely improve model capabilities, while building a data-production system that can continuously iterate.

You will work closely with the Foundation Model, Post-training, Agent, and scientific teams to build a continuous data loop: Data → Model → Agent → Scientific Workflow / Experiment → New Data.

Research Areas

• Scientific Data Engine

• Multimodal Scientific Data

• Scientific Data Mining & Knowledge Extraction

• Data Curation / Cleaning / Deduplication

• Synthetic Data Generation

• Agent Trajectory Data

• Reasoning Data

• Scientific Knowledge Extraction

• Data Quality & Data Evaluation

• Multimodal Data Alignment

• Data Pipeline & Large-scale Processing

• AI for Science

Responsibilities

• Scientific Data Engine: Design and build data-production systems for scientific foundation models and Scientific Agents, continuously generating high-quality training data from public scientific resources, papers, databases, research tools, agent trajectories, and other scientific environments.

• Scientific Data Extraction & Structuring: Process heterogeneous scientific information including paper text, tables, figures, molecules, sequences, structures, and experimental records; explore automated scientific knowledge extraction, structuring, and quality-control methods.

• Training Data Construction: Build datasets for different stages of pre-training, post-training, and agent learning, including multimodal data, reasoning data, trajectory data, preference data, and tool-use data.

• Data Quality & Evaluation: Build data-quality evaluation systems and study filtering, deduplication, contamination detection, difficulty estimation, data mixture, and related problems. Validate the real contribution of different data through model-training experiments.

• Data Flywheel: Work with Agent and scientific teams to turn successful paths, failures, tool calls, and experimental feedback from real research into new training data, creating an evolving scientific data flywheel.

• Data Infrastructure: Build scalable data pipelines for large-scale scientific data collection, cleaning, transformation, versioning, and training access.

What We Look For

• Bachelor’s degree or above in computer science, artificial intelligence, data science, mathematics, bioinformatics, computational biology, or a related STEM field.

• Experience with large-model training data, data engineering, data mining, information extraction, synthetic data, or related projects.

• Strong Python and data-processing / engineering skills.

• Experience with large-scale data pipelines and a practical understanding of the relationship between data quality, data distribution, and training outcomes.

• Ability to diagnose data problems from model-training results, not just perform ETL.

• Long-term interest in AI for Science and willingness to work with complex, heterogeneous, imperfect real-world scientific data.

Nice to Have

• Experience building pre-training or post-training data for LLMs / VLMs.

• Experience with synthetic data, reasoning data, or agent trajectory data.

• Experience with scientific literature mining, PDF / table / figure extraction, or knowledge extraction.

• Experience with scientific data in bioinformatics, cheminformatics, proteins, molecules, omics, or related areas.

• Experience with multimodal data alignment / representation.

• Experience with large-scale data pipelines, Ray / Spark, or distributed processing.

• Research experience in data-quality evaluation, data mixture, data ablation, or data scaling.

• Strong conference / journal publications, technical reports, or open-source projects.

We are especially interested in people who think deeply about one question:

What kinds of scientific data actually make models smarter, and how can we produce that data continuously and at scale?