Enabling a Major Pharmaceutical Firm with AI/ML Automation Across Hybrid Cloud Platforms

Shubham Khandelwal

Background

The R&D division of the pharmaceutical company manages multiple AI/ML models that accelerate the drug discovery process. With the increase in these models, there is also an increase in infrastructure and operational costs. This necessitates the right set of tools, platforms, and products to make the end-to-end process more efficient.

Problem statement

Creating solutions that address issues of scale, cost, efficiency, and complexity while providing a smooth user experience:

Establish robust infrastructure for scalable development and production pipelines, automating processes to efficiently manage a diverse project portfolio.
Address the limited availability of data scientists by implementing strategies to optimize resource allocation and effectively support project demands.
Streamline connections between complex data generators and consumers, enabling multi-modal optimizations to enhance operational efficiency.
Support the integration of AI/ML innovations into business processes seamlessly, empowering end-users with tools tailored to their needs and driving widespread adoption.

Solution

Highly resilient and scalable Intelligent Automation (IA) pipelines built on the company’s compliant, certified platforms.
- Enabling advanced AI/ML models at scale, catering to scientists globally and eliminating repetitive tasks with Intelligent Automation (IA).
- No-code solutions supporting the ML lifecycle, including model catalogues, hubs, and dashboards for driving insights.
Intuitive self-service and configurable low/no-code solutions that meet industry standards with just a few clicks.
- Enables easy reuse of complex data standardization pipelines. Connects with S3, NFS, DominoLabs, ABCD, Orbit, etc.
- Model explainability and tracking data/model drifts.
Company-specific and owned adaptor implementations connecting industry-leading technologies and tools.
- Enabling intricate customizations to enhance the day-to-day work experience of scientists, e.g., LiveDesign, 3Dx.
- Bringing in the best software engineering practices enabling reuse.
Adaptive hybrid solutions that can bring compute to where data resides or take data to compute efficiently.
- Connects imaging microscopes or huge Electronic Lab Notebooks (ELN) databases processing petabyte-scale data from on-premises or cloud sources and then optimizes the computation to bring optimal cost-to-value delivery.
- Enable state-of-the-art (SOTA) computing paradigms like quantum computing where applicable.
AI for AI: SOTA techniques to optimize AI models for training and inference use cases, bringing over 100% efficiency and thus reducing operation costs.
- Neural architectural graph optimizations.
- Hardware-specific software optimization for extreme performance/efficiency with multi-tenancy capabilities.
- SOTA AutoML capabilities: Neural Architectural Search, Reinforcement Learning, etc.

Conclusion

An optimized MLOps platform, designed to enhance cost efficiency across model training, inference, and deployment stages. It incorporates state-of-the-art (SOTA) open-source techniques to achieve these objectives.