Menú
Back to other articles

Designing the architecture of a sourcing platform assisted by AI— Part I

This article is the continuation of a series of posts covering applying AI and semantic search to tech recruitment. You can read them here:

Here I’ll start delving into the details of the system architecture of a sourcing platform empowered by AI. I hope this architecture can be used by companies as a reference to build up or re-structure their sourcing or AMS (Applicant Management System) platform.

The following are the primary functionalities of the platform:

  • Extracting key information from the applicant’s resume. Including (programming languages, technologies, frameworks, and education). That information can be used for keyphrase searches and candidate filtering. Also, label the candidate´s profile in the following categories: management, development, or QA. 💪.
  • Calculate the relevance of the candidate’s resume to a job post.
  • Generate resume summaries.
  • Generate cover letters for applicants.
  • Suggest salary ranges based on the Job description.
  • Generate transcripts and summaries from the audio of an interview call💪.
  • Detecting the applicant’s fluency in English 💪.
  • Automatically recommending candidates for job posts.
  • Send feedback to applicants 💪.

So, our platform has at least nine features, each one with its complexity. Also, we need to define the base architecture that will glue those functionalities together.

As that’s a lot of information to cover in a single article, I’ll focus in this one solely on the system design and on the following features:

  • Extracting key information from the applicant’s resume. Including (programming languages, technologies, frameworks, and education). That can be used for keyphrase searches and candidate filtering. Also, label the candidate´s profile in the following categories: management, development, or QA.
  • Calculate the relevance of the candidate’s resume to a job post.

The following is a simple mock-up (UI design is not one of my strengths 🙂 ) of the “Apply to job“ form. Almost all the features listed above are triggered just after the candidate presses the “Apply now” button.

The toolbox I propose in this article would allow:

– Car repair shops and insurance agents to automate their damage inspection process.

– Car washes can quickly generate an inspection of the cars they receive thus avoiding possible problems with car owners for surprise dents.

System design

Before delving into the details of the system design, we need to define some metrics about the platform, so we can use that information to design a sound architecture:

  • The platform receives, on average, one hundred (100) job applications daily per organization(you can assume we are building a multi-tenant system).
  • Resumes are in .docx and pdf format.
  • The average size of each resume is 500 Kb.
  • There are, on average, ten (10) users per organization.
  • 60% of users use the platform daily.
  • Daily active users (DAU)= 10 * 60% * 50 organizations = 300
  • Each user runs, on average, one (1) query each minute.
  • Each user spent, on average, 3 hours on the platform daily.
  • On average, the size of each query is 12 Kb.
  • On average, the size of each query response is 180 Kb.
  • The system can handle up to 50 organizations.
  • Plan for five years.

The previous information can help us size some of the following key elements:

  • The storage required to save the resumes can be calculated as follows: 500Kb * 100 resumes daily * 50 organizations * 365 days = 912.500.000 Kb. This is approx 912 Gigabytes per year or 4.56 Terabytes in five years.
  • Monthly inbound bandwidth, calculated from the average query payload : 12 Kb * 60 minutes * 3 hours* 300 DAU * 30 days = 19,4 Gigabytes.
  • Monthly outbound bandwidth, calculated from the average response payload: 180 Kb * 60 minutes* 3 hours * 300 DAU * 30 days = 291,6 Gigabytes.

Now, let’s make some architectural decisions based on the fact we want to design a performant, traceable, secure, and scalable platform:

  • We will implement asynchronous tasks for all the time-consuming features, for example, features that make inferences on AI models such as “Extracting key information from the applicant’s resume” or “Generating transcripts and summaries from the audio of interview calls.” That asynchronous tasks can be embodied in cloud functions or async workers.
  • We will define workflows to coordinate the correct execution of all the asynchronous tasks required to fulfill a business process; an example of a business process is “Register and Rank a candidate.” This process can then comprise the following tasks: “Extracting key information from the applicant’s resume” -> “Generating resume summaries”.
  • We will use a workflow engine or process coordinator like Temporal, to ensure a solid, scalable, traceable, and secure execution of our workflows.
  • We will save the candidate’s resume in blob storage such as google storage, an amazon bucket, a Digital Ocean’s object store, or even an on-premise database like SqlLite. So each process that uses the resume’s payload will receive the corresponding resume’s path as part of its execution parameter and need to get that payload from the storage.

The following is the list of stacks, frameworks, and tools we are using to build the solution:

In other words, CLIP can determine, given an image and a set of labels, which of that labels best describes the input image.

I encourage you to read the complete description of this amazing model here; I think it is good to know why CLIP is a game-changer in the visual classification field and what was the approach that OpenAI used to build it.

Implementing our first two async tasks

As we stated before, we want a performant and scalable platform. So using async processors or microservices orchestrated by a workflow engine is a good architectural decision.

Now is the time to start designing and developing the async task that will implement the following features:

1 — Extracting key information from the applicant’s resume. Including (programming languages, technologies, frameworks, and education). That can be used for keyphrase searches and candidate filtering. Also, label the candidate´s profile in the following categories: management, development, or QA.

2 — Ranking candidate’s resume to a job post.

Task 1 — Extracting key information from the applicant’s resume.

The command to trigger this task can be a JSON string as the following:

{
   "resume_path" : "path_to_resume_in_the_blog_storage",
   "job_post_id" : "e2c89a30-4a87-11ed-b878-0242ac120002",
   "candidate_id" : "f3ff2102-4a87-11ed-b878-0242ac120002",
   "organization_id" : "266edfc4-4a88-11ed-b878-0242ac120002",
   "workflow_id" : "108309444",
   "task-code" : "rs001"
}

We can expect the output of this task to be something like this:

{
  "candidate_id" : "f3ff2102-4a87-11ed-b878-0242ac120002",
  "workflow_id" : "108309444",
  "profile_details" : {
     "profile_categories": ['Developer'],
     "technical_skills" : ['Ethereum','Solidity', 'Golang', 'Node.js', 'Angular.js', 'React'],
     "education" : "I have M.Tech in Computer Engineering from Jaipur, Rajasthan Malaviya National Institute Of Technology Jaipur.",
     "relevance_score" : "0.734" 
 }
}

We have at least two alternatives to implementing the current functionality.

The first alternative is based on GPT-3. Here we extract the key information by sending a prompt with the candidate`s resume as context and some relevant questions and commands. Next, you can see an example of such a prompt:

Skills Strong CS fundamentals and problem solving Ethereum, Smart Contracts, Solidity skills Golang, Node, Angular, React Culturally fit for startup environment MongoDB, PostGresql, MySql Enthusiastic to learn new technologies AWS, Docker, Microservices Blockchain, Protocol, ConsensusEducation Details 
January 2014 M.Tech Computer Engineering Jaipur, Rajasthan Malaviya National Institute Of Technology JaipurBlockchain Engineer - XINFIN Orgnization
Skill Details 
MONGODB- Exprience - 16 months
CONTRACTS- Exprience - 12 months
MYSQL- Exprience - 9 months
AWS- Exprience - 6 months
PROBLEM SOLVING- Exprience - 6 monthsCompany Details 
company - XINFIN Orgnization
description - Xinfin is a global open source Hybrid Blockchain protocol.
Rolled out multiple blockchain based pilot projects on different use cases for various clients. Eg.
Tradefinex (Supply chain Management), Land Registry (Govt of MH), inFactor (Invoice Factoring)
Build a secure and scalable hosted wallet based on ERC 20 standards for XINFIN Network.
Working on production level blockchain use cases.
Technology: Ethereum Blockchain, Solidity, Smart Contracts, DAPPs, Nodejs
company - ORO WealthAnswer the following questions:1 - List the programming languages and frameworks you have worked with.  
2 - What did you study at school?  
3 - List the databases you have worked with. 
4 - Classify the candidate between the following categories: Developer, QA, DevOps, or Manager.
Answers:

Then we just have to parse the GPT-3 answer to provide the expected JSON response.

Answers:1
- I have worked with Ethereum, Solidity, MongoDB, AWS, Docker, and Microservices.
2 - I studied M.Tech Computer Engineering at Jaipur, Rajasthan Malaviya National Institute Of Technology Jaipur.
3 - I have worked with Ethereum, MongoDB, MySQL, and PostgreSQL.
4 - I would classify the candidate as a Developer.

Having the answer list from GTP-3, extracting the candidate category label is extremely easy; we just have to parse the fourth answer. In this case:

4 – I would classify the candidate as a Developer.

The second alternative to implementing this task is to train and use a custom Named Entity Recognition (NER) model based on Hugging Face Transformers to label programming languages, frameworks, and databases (more on that in another post).

Task 2 — Calculate the relevance of the candidate’s resume to a job post.

The idea here is to calculate how relevant is the candidate´s resume to the job post he is applying to.

We are using the SentenceTransformer library to extract the embeddings of both the job post and the candidate’s resume. Then we calculate the cosine similarity of both vector embeddings to see how similar they are.

from sentence_transformers import SentenceTransformer, util
model = SentenceTransformer('sentence-transformers/multi-qa-mpnet-base-dot-v1')

#Resume and job post are encoded by calling model.encode()
jobpost_emb = model.encode(jobpost)
resume_emb = model.encode(resume)

score = util.cos_sim(jobpost_emb, resume_emb)

We have reached the end of our adventure of designing the architecture of a sourcing platform assisted by AI. This is the first of a series of articles I plan to write to describe most of the technical challenges in creating a viable solution.

I hope this article has been helpful. Please add your comments if you have any questions or ideas about some other features you want to see on the platform.

Thanks for reading!

Stay tuned for more content about GPT-3, NLP, System design, and AI in general. I’m the CTO of an Engineering services company called Klever, you can visit our page and follow us on LinkedIn too.

Related Articles

Menú