Codecademy’s AI Resume Analyzer: The Job-Readiness Checker


Breaking into tech may be laborious with out steering. Many self-taught programmers and profession switchers have a troublesome time determining if their expertise are as much as par, and if not, what they should be taught to bridge the hole. You may’ve completed dozens of programs and even constructed just a few tasks, however is it sufficient to land a job? That’s why we’ve created our new AI resume analyzer: The job-readiness checker.  

Our new job-readiness checker, powered by GPT-4 (the newest mannequin from OpenAI, the corporate behind ChatGPT), helps consider how properly you meet the necessities for a given position based mostly in your expertise and expertise.  

If you’re logged in to Codecademy, you may entry the job-readiness checker beneath Assets in our navigation menu. After you have a job posting’s URL from LinkedIn, simply copy and paste the hyperlink alongside along with your resume (or simply copy and paste the complete job description from one other website). The job-readiness checker will parse by means of this info and the programs you’ve accomplished on Codecademy to generate a compatibility share that summarizes how properly you meet the necessities. Right here’s an instance of the way it can break down your compatibility and ability alignment with a given position. 

The job-readiness checker is designed to take the guess work out of your job search, so that you may be strategic about the place you ship your resume. “The job-readiness checker crystallizes how prepared you might be for a selected position,” says Owen Ou, Codecademy Senior Product Supervisor who helped spearhead its growth. With this instrument, we’ll additionally break down precisely which expertise you have already got and which of them you have to choose up. “We’re hoping that this empowers learners to know like, Oh, I’m solely 30% prepared,” he says. “Hopefully that may encourage some to undergo our content material sooner to speed up getting that final result.”   

Right here’s a peek behind the scenes at how our engineers used emergent AI expertise to construct our job-readiness instrument, and the way you should use the job-readiness checker in your individual job search.

Be taught one thing new without cost

The challenge: Create a instrument so learners can test in the event that they’re prepared to use for a job. 

If you’re studying to code so you will get a job in tech, it’s laborious to determine if you’re able to take the following step and enter the job market. Our learners have numerous instructional backgrounds and previous skilled experiences, so there’s no one-size-fits-all solution to consider job readiness. That’s why we thought this could be alternative to make use of generative AI.  

The largest duties concerned in growing the job-readiness checker included: 

  • Creating an OpenAPI endpoint 
  • Growing a brand new knowledge mannequin to save lots of customers’ delicate resume info 
  • Calibrating prompts to make sure ChatGPT offers the specified output 

Investigation and roadmapping 

Owen: “Earlier to AI, the one method {that a} learner would be capable to entry this type of functionality was to take a seat down with an actual human. A human advisor must overview what programs and tasks you’ve executed, how properly you’ve executed on every, and let you know manually [if you’re ready to apply] by reviewing the info that they’ve about you and looking out on the roles that you just’re focused on. 

I’d adopted developments in AI and neural networks for a few years. With the newest iteration popping out of OpenAI and the way simple and highly effective GPT-4 was, it appeared like we may begin tapping into the predictive capabilities a little bit bit extra. The challenge was mainly connecting a number of totally different dots: connecting the learner ache level; and connecting the enterprise curiosity to the expertise development that was growing. We thought this was an attention-grabbing drawback to work on. 

The primary half was only a tactical investigation. We had Jon, our tech lead, spend fairly a little bit of time initially simply taking part in round as a result of it’s a bleeding-edge expertise; nobody actually is aware of what it’s able to, and it adjustments each month. We had been simply tinkering and messing round with this expertise within the again finish in opposition to our speculation. This most likely took over a month, simply seeing what was succesful, and will it return a rating that roughly made sense based mostly on a Codecademy progress knowledge that we fed it? Directionally, we had been checking totally different parts of our imaginative and prescient and whether or not the present expertise may ship one thing. There’s like an enormous laundry record of instruments that we used — you’ll need to ask our Senior Software program Engineer Jon Sanders.” 


Jon: “That is the primary Codecademy function that makes use of ChatGPT, so we began by making a brand new micro service which comprises the endpoints for all our AI-related API calls (proper now simply OpenAI). This took a while to get proper, and we’re nonetheless iterating on it, nevertheless it’s good to have one service to trace all of our AI utilization. It permits us to see how a lot every function is spending on API calls, how they’re getting used, and we will add all of our numerous prompts there (amongst different issues). It’s good we constructed it early, as a result of now a number of groups are making use of it for upcoming AI-related options. 

As soon as we received the service operating, we began constructing the job-readiness checker’s frontend.  In some unspecified time in the future, I turned targeted on constructing the options as a substitute of modifying the prompts, so our workforce’s supervisor, Aditya Srinivasan and Curriculum Developer Melanie Williams, started work on calibrating the prompts to present us the specified outputs: correct scores of job-readiness and useful suggestions for what somebody ought to work on to change into extra appropriate with a given job posting. Aditya made a script to run many examples by means of the service at one time, so we will see how any immediate adjustments have an effect on a bunch of various instance learners. 

An important function of GPT-4 (additionally out there in GPT-3) is function_calling, which permits us to get predictably formatted JSON the way in which you’ll anticipate from a “regular” API.  Earlier than function_calling was launched, we had been having hassle making certain ChatGPT’s responses had been within the appropriate format, which made constructing the function almost unattainable. I think about nearly all the upcoming options from Codecademy that use ChatGPT may also make use of function_calling. It’s an important a part of the API for us.” 


Owen: “It’s a novel use case, and in consequence, we’ve needed to construct new playbooks. We’re simply utilizing rules and dealing by means of the muck, the mess, just like the low-level particulars to sort of obtain an goal. It’s not like we’re simply following a regular course of that different tasks have. It’s simply numerous on-the-fly drawback fixing, and utilizing assets and expertise throughout the firm that you just suppose have the uncooked expertise to assist deal with these issues. After which simply belief that it may be executed.” 

Jon: “We did numerous testing with GPT-3.5-turbo-0613 and GPT-4-0613 and located that we get way more correct and dependable outcomes with GPT-4.  We had hoped to make use of GPT-3.5 as a result of it’s sooner and cheaper, however the qualitative distinction in outcomes, for a immediate as difficult as we’re utilizing, was apparent. 

Sadly this implies our learners want to attend as much as 30 seconds for a job-report to get generated, so constructing a UI displays this with out being complicated was an attention-grabbing design problem that our designer, Mat Stevens, did an incredible job with.” 


Owen: “We had been all holding our breath to see if this could work. On daily basis we’d be like, Oh my god, may it do it? No. However then we’d attempt a piece round. If you’re doing stuff that hasn’t actually been executed earlier than, you already know that a few of your concepts simply aren’t going to be possible. However this one we had been each fortunate and well-timed. We knew the wave was coming, and we guess that this wave can be the proper wave to trip — and you would crash a little bit bit now and again.” 


Owen: “This one stood out in opposition to the opposite tasks I’ve labored on as a result of we didn’t even understand it was doable. We hadn’t seen this executed earlier than within the edtech business — that’s what’s distinctive about this challenge. We had been pioneering.” 

Jon: “I’d like so as to add a shout-out to Ahmed Abdallah, Workers Engineer — big snaps for constructing out numerous the AI-service, together with rate-limiting, infrastructure, and safety. His assist was important to the challenge.”


Leave a Comment