AI Fast-Track Methodology. Building Real-World AI Applications with Modern Frameworks. Azamat Sultanov

there’s one thing AI developers love, it’s building things. But here’s the kicker: most of the heavy lifting for common tasks has already been done. That’s why leveraging existing solutions is a cornerstone of the Fast-Track Methodology. The challenge isn’t building from scratch – it’s knowing where and how to look for the right tools and choosing the one that fits your needs.

      In this step, we’ll break down a full pipeline for finding and selecting the best pre-built libraries, frameworks, and models for your project. By the end, you’ll not only know what to use but also why it’s the right choice.

      Why Start With Existing Solutions?

      Here’s the golden rule: Don’t code what you can download. There’s no glory in reinventing the wheel when time, efficiency, and sanity are at stake. Pre-built solutions save you weeks (or months) and let you focus on building value, not wrestling with low-level details.

      For our AI-powered Digital Coach, we’re looking for tools that can handle:

      – Pose Detection: Identifying body key points in real time.

      – Repetition Counting: Analyzing sequences of movement.

      But how do you find these tools? Let’s dive into the search process.

      The Search Pipeline: How to Find the Right Tools

      Here’s a step-by-step guide to systematically search for existing libraries, models, and frameworks:

      1. Start with Google

      The simplest (and often most effective) way to begin is by Googling the task + “library” or “framework.” Here are example queries:

      – “pose estimation library”

      – “pose estimation python” – “real-time body tracking framework”

      Most of the time, this will lead you to:

      – Official websites of popular libraries.

      – GitHub repositories with large user bases – Blog posts or tutorials reviewing various options.

      Let’s see what pops up when we Google1 “pose estimation library” and “real time pose estimation python”:

      Fig 1: Google Search Results

      The search results include a mix of blog posts, official GitHub repositories, and even research-focused resources (we’ll explore those in more detail later). The types of entities you’ll see – blog posts, official repositories, research papers, or something else – will vary depending on the specific query you enter.

      2. Dive Into GitHub

      If Google doesn’t immediately land you on a goldmine, head over to GitHub. Use the platform’s search bar to query the task you’re looking for:

      – “pose estimation”

      – “AI exercise tracking” – “fitness computer vision”

      GitHub results often include repositories ranked by the number of stars. While stars aren’t everything, they’re a good indicator of popularity and community trust.

      Look for:

      – Repositories with a high number of stars (500+ is a great starting point). – Recent commits (active projects are less likely to break). – Clear documentation (so you can actually use the library).

      Let’s run the search query on GitHub2 for “pose estimation”, specify on Python and sort repositories by most stars:

      Fig 2: GitHub Search Results

      3. Explore Research Resources

      For cutting-edge tasks, you might need to look at academic papers. Most published research today comes with accompanying GitHub repositories. Great places to find this include:

      Papers with Code: 

      – This site pairs research papers with code implementations.

      – Use their search tool to find papers and projects related to your task. – Example query: “pose estimation.”

      arXiv:

      – While it’s primarily a preprint archive, many arXiv papers link directly to GitHub repos. – Bonus: You’ll often find links to model weights and pipeline instructions.

      Since these platforms are primarily geared toward finding research papers focused on specialized applications, let’s try searching for a more specific use case in pose estimation.

      For instance, consider “pose estimation for animals”.

      Let’s see what PapersWithCode delivers for this query:3:

      Fig 3: PapersWithCode Search Results

      These are just snippets of the results. Results also include benchmarks, libraries, datasets, starred papers, and repositories featuring the frameworks used.

      4. Evaluate Libraries Based on Key Criteria

      Once you’ve found a few potential solutions, it’s time to narrow them down. Use these criteria to decide which library fits your needs:

      – Ease of Use: Does the documentation make it easy to get started? Are there examples or tutorials?

      – Popularity: Look for libraries with a large user base and community support (stars on GitHub, forum discussions, etc.).