Low-code data science platforms: 3 things IT leaders should know

Organizations are increasingly turning to low-code machine-learning applications to address problems such as governance, time to market, and the talent shortage. Here’s what CIOs need to know
43 readers like this.
CIO_Big Data Decisions_1

Organizations across industries are turning to data and analytics to solve business challenges. A survey by New Vantage Partners found that 91 percent of enterprises have invested in AI. However, the same study found that just 26 percent of these firms have AI in widespread production.

Organizations are struggling to solve business challenges with AI. They find that building machine learning (ML) applications takes time and requires expensive maintenance and talent that’s in short supply. Leaders say that over 70% of data science projects report minimal or zero business impact.

Here’s how low-code ML platforms can help tackle these challenges.

Low-code is a software development approach that leverages a visual user interface to create applications instead of traditional hand-coding. For decades, developers built applications by writing thousands of lines of code from scratch, often round-the-clock.

[ Also read 3 automation trends happening right now. ]

Building software solutions using low-code falls somewhere in the continuum between programming from scratch and buying off-the-shelf. It brings the best of both worlds by balancing flexibility and time-to-market.

A low-code development platform (LCDP) is considered quicker to build, economical to maintain, and developer-friendly because of its visual approach.

Low-code tools empower enterprises by democratizing software development. Today, anyone with a business interest and basic technology skills can build an app using low-code technology. According to Gartner, by 2024, more than 65 percent of all app development will be on low code. Globally, the low-code market is projected to reach $187 billion by 2030.

Can low-code accelerate AI solutions?

Globally, the ML market is projected to reach $209 B by 2029 at a compounded growth rate of 38.8 percent. Machine Learning Operations (MLOps) is a set of practices that came into the limelight in recent years. By streamlining software operations and simplifying collaboration amongst data science and development teams, MLOps helps build production-grade AI solutions.

In essence, MLOps delivers the goods through three practices: Continuous Integration (CI), Continuous Delivery (CD), and Continuous Training (CT).

CI deals with the automatic building and integration of code from multiple contributors into a single application. CD is the practice of continuously and predictably delivering quality products to production. CT ensures monitoring and retraining of the ML model using new data when model performance begins to dip.

Why do organizations struggle to build, scale, and deliver value with ML? There are three key challenges:

  • Long cycle time: Building robust AI models at enterprise scale takes time. 80% of companies say it took them six months to productionize an AI model.
  • Model drift: With continuous changes in the external market, business dynamics, and foundational data, models tend to go stale rapidly. Model drift leads to a drop in accuracy and poor business decisions.
  • Talent shortage: Data science practitioners who can solve business challenges by applying AI are in short supply. VentureBeat opines that shortage of skills is one of the major reasons behind slow AI adoption.

A low-code methodology addresses these challenges by bringing a visual, automated approach to MLOps. It helps accelerate go-to-market, enables efficient model maintenance, and democratizes data science development by lowering skill barriers.

[ Also read 3 essentials for a low- and no-code application development strategy ]

How low-code data science platforms address the MLOps challenges

There are three ways a low-code platform tackles the roadblocks most data science teams face:

1. Quicker time to market

Low-code platforms can speed up development by offering reusable components needed throughout the ML lifecycle – data connectors, data handlers, backend/frontend development modules, ML algorithms, visualization widgets, and administration and security modules.

By providing a ready-to-use library in a drag-and-drop approach, it allows developers to build and bug-fix rapidly. This makes it easy for the data science team to collaborate, iterate, and optimize until the business challenge is addressed.

2. Easier model maintenance and improved governance

When trained ML algorithms risk going stale even before they go live, low-code tooling offers efficient ways to keep them refreshed. They make it easy to monitor models continuously, detect model degradation, and automatically take action through centralized governance.

Low-code ML platforms help detect model drift by flagging trigger-based alerts. They provide mechanisms to retrain models at defined thresholds and dynamically swap out models based on performance. By operationalizing the MLOps practices of CI-CD-CT, low-code ML platforms help tackle model maintenance issues.

3. Bridging the skill gap

Every organization, large or small struggles to find, engage, and retain data science talent. By offering an intuitive drag-and-drop interface, low-code platforms crash the barriers to data science development.

With low-code platforms, it is easy to retrain an in-house software development team for ML needs. Reusable components in a repeatable workflow make it less cumbersome to retain knowledge about AI applications or maintain them with new hires. This translates to lower costs in training and ML development.

Cold-chain logistics provider United States Cold Storage (USCS) aimed to reduce warehouse turn times to improve customer experience and avoid hefty penalties. They embraced low-code ML platforms to develop an automated appointment scheduler. USCS called in data science experts to identify the root cause of this delay – a manual appointment scheduling system. They used low-code tooling to build a predictive scheduler within a quarter.

After piloting the solution at one warehouse, it was productionized across 26 warehouses across the U.S. This led to a 16 percent reduction in warehouse turn times and savings of $300K over a quarter. The low-code platform allowed the USCS team to rapidly build and deploy the solution at scale with minimal demands on their technology personnel.

Embrace low-code data science platforms to lower the total cost of ownership

Organizations often embark on low-code data science development by evaluating technology platforms. This can be disastrous. The best place to begin the low-code journey is by starting with organizational priorities and understanding business needs in the short and long term.

Evaluate low-code platforms by checking alignment with the organization’s technology strategy, architecture, and roadmap. Make your choice based on the total cost of ownership by factoring in the spending on tools, people, and process changes across the build and maintenance phases.

[ Want best practices for AI workloads? Get the eBook: Top considerations for building a production-ready AI/ML environment. ]

Ganes Kesari is an entrepreneur, AI thought leader, author, and TEDx speaker. He co-founded Gramener, where he heads Data Science Advisory and Innovation. He advises executives of large organizations on data-driven decision making.