Skip to content

Designing Solutions for ML Systems

Goku Mohandas
· ·

Designing a solution with constraints.
Repository ยท Video

๐Ÿ“ฌ  Receive new lessons straight to your inbox (once a month) and join 20K+ developers in learning how to responsibly deliver value with ML.


Once we've identified our main objective, we can hypothesize solutions using a three-step process: visualize, understand and design.


Visualize an ideal solution to our problem without factoring in constraints. It may seem like a waste of time to think freely without constraints, but it's a chance to think creatively.

  • most creative solutions start from a blank slate
  • void of the bias from previous approaches


Understand how the problem is currently being solved (if at all) and the how and why things are currently done the way they are.

  • prevents us from reinventing the wheel
  • gives insight into processes and signals
  • opportunity to question everything


Design from our ideal solution while factoring in constraints.

Automate or augment?

  • be wary of completely removing the user
  • transition from augment to automate as trust grows

UX constraints

  • privacy, personalization, property
  • dictate the components of our solution

Tech constraints

  • data, time, performance, cost, interpretability, latency
  • dictate the complexity of our solutions


The main goal here is to think like a problem solver, as opposed to a naive model fitter.

โŒ Model fitter โœ… Problem solver
naively maps a set of inputs to outputs knows which set of inputs and outputs are worth mapping
obsesses on methods (models, SOTA, single metric, etc.) focuses on product (objective, constraints, evaluation, etc.)
fitting methods are ephemeral foundational mental models are enduring


Our main objective is to allow users to discover the precise resource.

  1. Visualize The ideal solution would be to ensure that all projects have the proper metadata (tags) so users can discover them.

  2. Understand So far users search for projects using tags. It's important to note here that there are other available signals about each project such as the title, description, details, etc. which are not used in the search process. So this is good time to ask why we only rely on tags as opposed to the full text available? Tags are added by the project's author and they represent core concepts that the project covers. This is more meaningful than keywords found in the project's details because the presence of a keyword does not necessarily signify that it's a core concept. Additionally, many tags are inferred and don't explicitly exist in the metadata such as natural-language-processing, etc. But what we will do is use the other text metadata to determine relevant tags.

  3. Design So we would like all projects to have the appropriate tags and we have the necessary information (title, description, etc.) to meet that requirement.

    • Augment vs. automate: We will decide to augment the user with our solution as opposed to automating the process of adding tags. This is so we can ensure that our suggested tags are in fact relevant to the project and this gives us an opportunity to use the author's decision as feedback for our solution.
    • UX constraints: We also want to keep an eye on the number of tags we suggest because suggesting too many will clutter the screen and overwhelm the user.
    • Tech constraints: we will need to maintain low latency (>100ms) when providing our generated tags since authors complete the entire process within a minute.
UX of our hypothetical solution


For the purpose of this course, we're going to develop a solution that involves applied machine learning in production. However, we would also do A/B testing with other approaches such as simply altering the process where users add tags to projects. Currently, the tagging process involves adding tags into an input box but what if we could separate the process into sections like frameworks, tasks, algorithms, etc. to guide the user to add relevant tags. This is a simple solution that needs to be tested against other approaches for effectiveness.


To cite this lesson, please use:

    title  = "Solution - Made With ML",
    author = "Goku Mohandas",
    url    = ""
    year   = "2021",