MIDS Capstone Project Spring 2025

AI Career Adviser

Team members
Bhaskar Jyoti Borah
Karen Orozco
Xin Chen
William Li
Torrey Trahanovsky

AI Career Adviser combines cutting-edge AI technology with deep industry insights to help you navigate your professional growth. Whether you're looking to advance in your current field, pivot to a new industry, or simply maximize your potential, our personalized guidance is designed to help you succeed in today's competitive job market.

Please check us out at: https://www.aicareeradviser.us/

Problem & Motivation

Currently, today's AI tools and especially today's career advisors are relying on outdated methodologies for providing skill-gap analysis to those early in their careers. The advice tends to be generic, not informed by real-time data intelligence, and does not augment LLMs or leverage agentic AI in the process. 

Additionally, the skills that are being learned today are evolving, with 70% of the skills used in most jobs changing, per research conducted by LinkedIn. Additionally, In the U.S., 12.4 million people are actively seeking jobs each month. The career and education software market is currently valued at $28 billion, with annual growth projected between 8% and 15%. Globally, EdTech is growing as an industry as well, driven by increasing demand for digital learning and AI-powered tools.

AI Career Advisor seeks to bring accessible, up-to-date, and privacy respecting career advising to those early in their career, using a proprietary dataset and algorithm. The market for career education counseling is estimated to grow to $5.64B by 2033 with few AI players, especially when compared with AI Career Advisor. 

The product has 2 unique selling points that solve pain-points voiced repeatedly to this capstone group during initial concepting research: 1) It’s skill-oriented and industry aware, giving users the quick ability to identify skill-gaps and address them based on the latest trends. 2) It’s LLM-centered with privacy as a priority, ensuring respect for our users' data and a less bias-prone advising experience. 

The job market is evolving quickly, and those early in their careers want to understand what skills are in-demand, and how they can work on refining these skills to obtain the most desirable/trending jobs on the market. For the purposes of the capstone, the job in-focus is Data Scientist.

Our MVP is designed to help users seamlessly transition into Data Scientist roles across three key industries: Technology & Software, Healthcare & Pharma, and Finance & Banking.

Data Source & Data Science Approach

Data Sources:

Approach:

Skills Generation and Standardization - To understand what skills a Data Scientist truly needs—by industry, not in general - we built a novel dataset by distilling millions of job postings, enhanced with LLM analysis, and real-time job signals. Using GPT-4o, we extracted job-specific skills, then categorized them into Domain, Technical, Soft, and Academic groups to build tailored role profiles. Finally, we enriched these profiles with real-time data to ensure they’re current and market-aligned.

Skill Gap Generation - Utilized the skills generated above + enhanced further with industry/domain expertise to create a final usable dataset of Data Scientist roles skills in 4x industries (General, Finance, Health, and Tech). These role skills, compared with a user’s resume skills would produce "gaps" in the skills of users.

Recommendations - Utilized the EDX (augmented with LLM for descriptions), Multi-Course (Skillshare courses specifically), and Coursera dataset to create recommendations. Additionally, augmented this with the Top 50 conferences.

RAG/LLM Approach - For any remaining recommendations, used LLM with these datasets (prompted) to ensure the LLM would suggest recommendations informed but not limited to these datasets. Tested with empty resumes with hHuman review to ensure all recommendations met expectations.

The overall architecture begins with Step 1 on the left, where we process a large volume of job posting data to generate role-specific skills. On the right, Step 2 enables users to upload their resumes. These inputs are then passed to the central AI Service, in Step 3, which orchestrates the core logic, including resume skill extraction, skill gap analysis by comparing the resume skills and role skills, and generating personalized recommendations.

The entire pipeline is fully deployed and operational on Amazon AWS, ensuring scalability and reliability.

We approached the Role Skills Generation mentioned in Step 1 as below:

As the first step, we used GPT-4o, fine-tuned with multiple custom prompts, to extract job-specific skills from millions of job descriptions.

In step 2, we consolidated all extracted skills and categorized them into four key groups: Domain, Technical, Soft and Academic Skills.

This enabled us to build tailored skill profiles for Data Scientist roles.

Finally, in Step 3, we enriched these role skills using live job posting data, which helped us create enhanced, current, market-aligned skill sets.

Once we have  our Data Scientist role skills, we worked on the skill gap analysis, model and personalized recommendations. We approached this using our base and enhanced model respectively.

Our base pipeline starts with a large language model (GPT-4o) that extracts skills from resumes and matches them against distilled role skills generated from job postings. This forms the foundation for three core components:

  • Skill Generation
  • Gap Analysis
  • Career Recommendations

To significantly enhance accuracy, and personalization, we introduced three major model improvements to enhance our base model:

  • Adversarial Dual Model with Prompt Tuning: We fine-tuned prompts across both resume and job skills streams using a dual-model approach. Here we leveraged and pitted two different large language models against each other, compared the prompts that they generated and utilized the better ones.
  • LLM + SkillBERT Embedding Integration: Enhanced gap analysis between the role and resume skills, by integrating SkillBERT embeddings with LLM outputs. This resulted in more accurate skill gap results.
  • LLM + RAG (Retrieval-Augmented Generation) Search: We strengthened our recommendation process by using RAG search. This enabled us to provide accurate personalized recommendations based on each user’s skill gap.

These enhancements showed a marked improvement in our enhanced model vs our base model, making it more scalable, intelligent and personalized.

 

Our platform is built with fully automated ML operations using GitHub Actions and AWS Services, enabling rapid development, testing, iteration, continuous deployment, and seamless scaling.

Evaluation

2 types of Evaluation were conducted:

  1. Quantitative - Model Evaluation.
  2. Qualitative - User Evaluation.

Key Learnings & Impact

Through user research and successfully building this product, and surveying it via those early in career, our team demonstrated the importance of integrating cutting-edge AI technology with deep industry insights to provide personalized, actionable career guidance.

One of the key learnings is the necessity of skill-oriented and industry-aware advising. By focusing on the latest trends and real-time data, the AI Career Adviser has the potential to quickly identify skill gaps and offer targeted recommendations, addressing a major pain point for early career professionals.

Another key feature is the project's ability to provide privacy to users (based on feedback in the semester from the privacy audit) while also reducing bias in advising to ensure a respectful and equitable experience for users.

Our Achievement can be summarized as successfully taking a concept, surveying and interviewing users to build the correct product, then once built, obtaining feedback to confirm the product meets the users aspired solution. This was successfully achieved in just a semester.

The key impact was:

  • This team developed a fully functional AI career adviser web application designed to assist job seekers by leveraging AI to identify skill gaps and generate personalized improvement suggestions.
  • Implemented industry selection and comparison features, allowing users to select different target industries and compare their skill strengths and gaps.
  • Conducted two rounds of feedback sessions with job seekers to enhance the application's features and user experience.
  • Engaged in additional SME chats with professional career advisers to further refine the product.

Technical Challenges Overcome / Learnings from Overcoming

  • Efficiently processed large volumes of job postings.
  • Prioritized critical skills for better user guidance.
  • Achieved fully automated MLOps for seamless operations.
  • Developed a comprehensive role skill dataset.
  • Provided actionable personalized recommendations to users.

Acknowledgements

Additionally, for this project, if time allowed, we would seek to further refine the recommendation dataset and functionality of the website, as enriched recommendation sources and variety was a key user feedback & opportunities noticed. 

We also would like to implement email reminders for personalized to-dos to increase user retention and completion of their personalized plan.

Finally, if time allowed, we would have liked to offer “Job Matching” functionality, in which the tool could alert users about jobs matching based on their best matching skills using real-time job data.

More Information

Official Website - HomePage - AI Career Advisor
Final Presentation Slides
Ai Career Advisor
Ai Career Advisor

Video

Last updated: April 21, 2025