Build Custom AI Recruiting Workflows

Most firms start with their ATS's built-in filters, like in Greenhouse or Lever. These are simple keyword matches. A filter for "Python" and "FastAPI" will find resumes that list both, but it cannot rank them. It treats a candidate with 10 years of Python experience and a 3-month FastAPI project the same as an expert. This creates a high volume of low-quality matches that a human still has to sort through.

Consider a 15-person tech recruiting firm sourcing for a startup client. The client needs a "senior Python developer with FastAPI experience who has worked at a company under 50 people." The ATS filter for "FastAPI" returns 80 candidates. A recruiter must manually open each resume or LinkedIn profile to check their work history for company size. This manual check for a single role takes 2-3 hours and has to be repeated every week.

This is a design limitation. These platforms are databases with search, not intelligent matching systems. They cannot combine signals, score candidates on a spectrum, or personalize outreach based on a candidate's specific project history. The logic is rigid: a candidate either matches a keyword or they do not. This binary approach breaks down for nuanced, high-value roles.

Syntora would start by auditing your existing applicant data and ATS integration points, whether you use Lever, Greenhouse, or Ashby. This initial discovery phase would allow us to understand your specific data structure and needs. The approach would involve connecting to your ATS API to securely pull relevant applicant data. For efficient data retrieval, we typically utilize Python's httpx library for async requests. Each resume PDF or DOCX would be parsed using PyMuPDF to extract raw text, which we would then store in a structured Supabase Postgres database. This process creates a clean, normalized dataset essential for effective model training. We have experience building similar document processing pipelines using Claude API for sensitive financial documents, and the same robust pattern applies to recruiting documents.

For each open role, the system would utilize your job description and a set of "gold standard" resumes you provide to inform model behavior via the Claude API. The core matching logic would be built in a FastAPI application, designed to generate a comprehensive set of features for each new candidate and rank them based on alignment with the role's requirements.

For highly matched candidates, the system would be capable of generating personalized outreach emails. This functionality would leverage the Claude API to highlight specific candidate alignments discovered during the screening process, for example, noting "your experience leading FastAPI migrations aligns directly with our client's current project." Structured logging via structlog would be implemented to track every generated message, enabling quality control and continuous improvement.

The entire system would be deployed as a set of serverless functions on AWS Lambda, triggered by webhooks from your ATS when a new candidate applies. The process would write the match score and a suggested outreach email back to a custom field in your ATS. Hosting costs for such an architecture are typically under $50/month. A simple front-end dashboard built with Vercel would provide system status and insights. Typical build timelines for an initial system of this complexity range from 8 to 12 weeks, depending on the number of integration points and the cleanliness of existing data. The client would need to provide access to their ATS API, historical applicant data, and a selection of example "gold standard" resumes for each role. Deliverables would include the deployed AI system, source code, and comprehensive documentation.