KLEROS AI

Inspiration

The current recruitment landscape is flawed. [cite_start]Hiring decisions are frequently influenced by unconscious bias, causing the unfair exclusion of capable candidates[cite: 11]. [cite_start]Decisions are often based on gender, religion, region, or educational background rather than actual capability[cite: 12]. [cite_start]We noticed that talented individuals from Tier-2 and Tier-3 institutions are often overlooked simply because they do not fit a traditional pedigree profile[cite: 13].

[cite_start]We built Kleros AI because we believe that innovation relies on fairness[cite: 14]. [cite_start]Our goal is to create a system where talent meets unbiased opportunity, allowing companies to build balanced and high-performing teams based purely on merit[cite: 2, 15].

What it does

Kleros AI is an end-to-end platform designed to eliminate bias from the initial screening process. [cite_start]It operates through a streamlined three-step workflow[cite: 34]:

[cite_start]Resume Anonymizer: The system automatically ingests resumes and strips away all personal identifiers including name, gender, college name, and location[cite: 26]. [cite_start]This shifts the focus entirely to relevant experience and qualifications[cite: 30].
[cite_start]AI Skill Fit Scoring: Using a private AI model, we evaluate the anonymized data against role criteria to generate an objective ranking[cite: 27]. [cite_start]This provides a bias-free score of capability and potential[cite: 31].
[cite_start]Fair Interview Coach: To ensure the interview phase is also equitable, we provide candidates with access to structured interview practice and feedback[cite: 28, 38].

How we built it

[cite_start]We architected Kleros AI using a modern microservices approach to ensure scalability and performance[cite: 41]:

[cite_start]Frontend: We built a responsive and interactive UI using React.js and Next.js to ensure a smooth user experience[cite: 42].
[cite_start]Backend: We utilized a hybrid backend using FastAPI for high-performance ML inference and Flask for general application logic[cite: 42].
AI & ML: This is the core of our engine. [cite_start]We used Tesseract and Google Vision API for OCR to extract text from various resume formats[cite: 43]. [cite_start]We implemented SpaCy and HuggingFace transformers to perform Named Entity Recognition for the anonymization process and semantic analysis for skill matching[cite: 43].
[cite_start]Database: We used PostgreSQL for relational data and MongoDB for unstructured resume data[cite: 42].
[cite_start]DevOps: The application is containerized using Docker and deployed on AWS[cite: 43].

Challenges we faced

One of the biggest technical challenges was accurately distinguishing between identifying information and relevant information. For example, removing a college name while keeping the degree title intact required fine-tuning our NLP models. Additionally, parsing PDF resumes with complex layouts using OCR presented significant data cleaning hurdles.

Accomplishments that we're proud of

We are most proud of our Skill Fit Algorithm. We managed to create a mathematical representation of a candidate's suitability that minimizes demographic influence.

If we define a candidate's capability as $ C $ and the role requirements as $ R $, our model strives to maximize the match score $ S $ while minimizing the bias coefficient $ \beta $:

$$S_{final} = \frac{\sum_{i=1}^{n} (w_i \cdot skill_i)}{1 + \beta_{demographics}}$$

In Kleros AI, our goal is for $ \beta_{demographics} \to 0 $, making the score purely a function of weighted skills.

What we learned

We learned that fairness is not just a social concept but an engineering challenge. We gained deep insights into how AI models can be tuned to recognize context without hallucinating data. We also learned how to orchestrate multiple services into a synchronous pipeline, linking OCR, NLP, and Generative AI seamlessly.

What's next for Kleros AI

Currently, we offer interview practice, but we plan to expand this into a real-time copilot for interviewers that flags biased questions as they happen. We also aim to integrate with major Applicant Tracking Systems to bring unbiased hiring to enterprise workflows.

Change starts with fairness. [cite_start]Fairness starts with us. [cite: 55]