When setting up MLOps at your company, you need to be sure that the people in your company are well versed in the stack, the stack is as good as it comes, and the work achieved is scalable and deployable. This three-part is a step-by-step guide to help make your own MLOps Setup as successful as possible.
An MLOps team would help your company meet its goals in a much better way through the help of its members. These roles are responsible for analyzing and developing machine learning models. Therefore, the MLOps team needs people who know how to code, know what different data points mean, and understand how human language will impact your startup's product.
We'll discuss setting up an MLOps team with the right talent, support, and training. We'll go over best practices from hiring to challenging and tech-savvy development and data science teams who want to work in an ultra-competitive environment that doesn't have a policy, process, or technical requirements like your typical tech industry does!
Do you need an MLOps Team at your Startup?
Any investment comes with a need, and that is no different from hiring a team of personnel that will deal with your daily day-to-day data, ML, Development, Deployment, and Monitoring for your cause.
Let us before diving into the analysis of whether your startup needs an MLOps team or not, let us see the stages your startup may be in right now:
The Pre-Seed stage is the phase of analysis, a crucial stage where you detect a real problem in a niche market. Some questions that need answers at this stage are:
Is your solution a good solution for the problem you are trying to solve?
Does your solution already exist in some way or another?
Is your solution radical for the market, and how will it be perceived?
Seed Stage: the phase of validation, where you have to develop a solid business model that is ready for confirmation from the investors on the shoulders of some prototypes. Some questions that need answers at this stage are:
How will rejection or non-validation be dealt with?
Is the product at a Minimum Viable Product stage for the investors?
How far are you towards the materialization of your startup?
Series A: the phase of an established track record, a stage where a startup has established a track record (a conventional revenue stream and customer base). Typically raising around $2 Mil to $15 Mil, you are ready to raise substantial money to expand your product and user base.
Some questions that need answers at this stage are:
What should be the desired funding stream, i.e., Angel Investors, Venture Capital?
What is going to be the political process of the working of the company?
Is crowdfunding an option for generating capital?
This stage serves as a reality check for most companies, as only fewer than 10% of the seed-funded companies will go on to raise Series A funds.
Series B: Beyond the development stage, this financing is solely focused on growing market penetration. Companies who have gone through the seed and Series A investment rounds have already built a client base, which gives them the confidence to prepare for larger-scale success.
Once we have identified the stage of the startup, we need to answer the following questions that are going to dictate the amount of data-oriented staff that is required for the startup:
What are the features necessary for ML design?
What part of your business can be modeled into a data science problem?
Do we have the data or pipelines required for it? (You are on the right website if you are looking for ML Pipelines 😉)
What is the quality of data that you are collecting?
How are you collecting it?
How well is the data annotated?
How will the model be presented to the user?
Do we have appropriate monitoring metrics for the model in production?
Once you have identified the current stage of your business and formulated these questions, let us start looking at the team you need to hire.
* * * * *
Roles in MLOps team
When assembling the data scientists for your startup at any stage, we need to remember the 7 Questions mentioned above. Answering these questions falls onto the shoulders of the following roles that need to be hired through a meticulous process of scouting and training.
Data Analyst
Data analysts work closely with product managers and business teams to conclude from the information gathered about users through surveys and other means. The main goal of the job is to figure out how different insights relate to each other and do the necessary statistical analysis on them.
Skills to look out for: Descriptive Analysis, Comprehensive Analysis.
Tech Stack: SQL, Excel, Tableau.
How many do you need: 2
Data Engineer
Data engineers ensure that the infrastructure to collect, transform/process, and store data is well built. They also decide how and in what format data goes into the Machine Learning pipeline. They are also in charge of the Transform Load jobs, which involve taking data from a source, processing it, and storing it in data warehouses. Data engineer takes the requirements of the stakeholders and builds a solution to fit their needs after understanding all the tradeoffs.
Skills to look out for: Distributed System Fundamentals, Data Structures, Algorithms.
Tech Stack: Spark, Hadoop.
How many do you need: 1
Data Scientist
Data Scientists are the most common and well-known of all the professions, and they are in charge of creating something useful from the data that has been collected and processed. They are responsible for analyzing, processing, and interpreting data and developing Machine Learning models. They use the infrastructure built by the data engineer to meet their goals and can also be responsible for building metrics for production monitoring.
Skills to look out for: Knowledge of the subject, Communication, and Interpersonal skills to deliver the business requirements.
Tech Stack: SQL, Python/R, and Machine Learning
How many do you need: 2
Research/Applied Scientist
Companies that want to focus on innovation, cutting-edge technology, and never-before-seen solutions should pay more attention to this job. The staff here tries to build new algorithms to conquer the area. However, they are not required to adapt their work to production.
Skills to look out for: Expert in sub-fields, i.e., NLP, CV, Statistics, Time-Series.
Tech Stack: SQL, Python/R, and Machine Learning
How many do you need: 3-4 (Depending on need and budget)
ML Engineers
The position is sometimes conflated with that of a Data Scientist, although it differs in that it focuses on the infrastructure and software development associated with model creation. Their task includes developing tools for upgrading models and designing user-friendly prediction interfaces. They collaborate closely with data scientists and use their models. Their responsibilities include utilizing APIs to provide usable endpoints for both developers and consumers.
Skills to look out for: Experience with Development paired with Deployment
Tech Stack: FastAPI, Docker, Kubernetes, Machine Learning
How many do you need: 2 (Balancing the Data Scientists)
Developers
Developers are the last cog in the system that will eventually connect your complete machine learning pipeline to your main application. Their job includes the seamless integration of the whole pipeline from Data Ingestion to Output Generation to the final product. They can be the on-call for monitoring and acting as general-purpose engineers trying to fill gaps where needed.
Skills to look out for: Back-end Developers, API Generation, Integration.
Tech Stack: Rest, AWS Lambda, AppSync, etc.
How many do you need: 2-3
* * * * *
Need help finding such talented people for your startup? How about putting an opening NimbleBox's very own hub of talented MLOps enthusiasts, Everything MlOps!
Congratulations! You have your MLOps team at your budding startup at your disposal, fulfilling your MLOps needs. This process brings you closer to setting up a functioning MLOps department at your startup. Catch the next part in the series in the upcoming week, where we talk about the intricacies of "Build vs. Buy"!
Written By
Aryan Kargwal
Data Evangelist