Reduce Hiring Bias with a Custom AI Screening System

Most small companies rely on their ATS for initial screening, but tools like Greenhouse and Lever offer primitive keyword matching, not true AI. This approach is easily biased. A filter for 'MIT' or 'Stanford' might screen out qualified candidates from state universities. Relying on years of experience can discriminate against younger applicants with equivalent skills.

A 30-person software firm we worked with was hiring a senior engineer. Their recruiter set up a filter in their ATS to find candidates with 10+ years of experience who worked at FAANG companies. This simple rule-based filter immediately disqualified 80% of their applicant pool, including several high-potential candidates from successful startups and open-source contributors who were ultimately better fits. Manual review is no better; studies show reviewers spend just 7 seconds on a resume, relying on pattern-matching that reflects their own biases.

These built-in ATS features create a false sense of objectivity. They automate a flawed manual process, encoding existing biases into the workflow. Because the logic is simple keywords and boolean flags, it cannot understand context, skill equivalency, or a candidate's trajectory. This leads to a homogenous pipeline and missed opportunities.

Syntora's approach to reducing bias in hiring would begin with a discovery phase. We would audit your existing hiring process and data within your Applicant Tracking System (ATS) to identify key data points and potential areas for bias. We would then work with your team to define specific, objective criteria for candidate evaluation and establish a project scope.

The technical implementation would involve connecting to your ATS API to pull historical applicant data, including resumes and hiring outcomes. We have built document processing pipelines using Claude API for financial documents, and the same pattern applies to parsing unstructured text from resumes and cover letters. The Claude API would parse this text into a standardized JSON object, extracting features such as technical skills, certifications, and years of experience with specific tools. This creates a clean dataset for model training.

Using Python and scikit-learn, we would train a gradient boosting model on this structured data. The model would learn the statistical relationship between skills on a resume and your historical hiring decisions (e.g., advanced to interview, hired). Crucially, we would explicitly exclude demographic proxies like names, locations, and university prestige from the feature set, focusing the model purely on demonstrated abilities and experience.

The trained model would be packaged as a lightweight API using FastAPI and deployed to a serverless function on AWS Lambda. When a new candidate applies through your ATS, a webhook would trigger the function. The API would ingest the resume, generate a score, and write that score plus a short summary of the candidate's top skills back into a custom field within your ATS.

The delivered system would provide a ranked shortlist, not an automated decision, ensuring a human recruiter still makes the final choice. We would also build a simple dashboard using Streamlit that tracks key fairness metrics, such as the pass-through rate for different demographic groups, to monitor model performance. Typical build timelines for this complexity range from 6 to 10 weeks, depending on data availability and client integration needs. Your team would need to provide access to ATS data, clearly defined hiring criteria, and ongoing feedback during the training and validation phases.