FlexiHire
Spearheaded a dual-sided intelligent matching recommender system, deployed as a production-grade FastAPI service with low latency via in-memory caching. It connects job seekers and companies through content-based filtering. Formulated a hybrid search strategy in Qdrant backed by a fine-tuned Sentence Transformer, significantly boosting MRR and precision over the baseline. Developed a multi-agent AI pipeline utilizing CrewAI and GPT-3.5-turbo to automatically extract, validate, and clean skills from raw CVs and job descriptions.
Technologies Used
Problem Statement
In the modern recruitment landscape, companies lose millions of dollars to prolonged hiring cycles and misaligned talent acquisition. Traditional Applicant Tracking Systems rely on rigid keyword matching, resulting in high false-negative rates where exceptional candidates are overlooked. Job seekers experience a "black hole" effect, submitting applications without feedback. The business problem is a severe inefficiency in matching human capital to organizational needs, leading to lost productivity and inflated recruitment costs.
Solution
FlexiHire acts as a strategic talent acquisition solution by implementing an intelligent, dual-sided recommendation engine. By utilizing hybrid search with Qdrant and fine-tuned Sentence Transformers, it understands the semantic context of both job descriptions and candidate CVs beyond simple keywords. The integration of a multi-agent AI pipeline using CrewAI automatically extracts, validates, and cleans skills, standardizing the data layer. This drastically reduces time-to-hire, increases candidate-role fit precision, and ultimately saves businesses significant human resource expenditure while providing a superior experience for applicants.
Key Features
Dual-sided intelligent matching recommender system
Production-grade FastAPI service with in-memory caching for low latency
Hybrid search strategy in Qdrant with fine-tuned Sentence Transformers
Multi-agent AI pipeline using CrewAI for skill extraction and validation
Automated cleaning of skills from raw CVs and job descriptions
Engineering Challenges
Aligning the semantic space of diverse job descriptions and complex CV formats
Ensuring the multi-agent extraction loop maintained high fidelity without hallucinations
Balancing retrieval latency with high-accuracy hybrid search constraints
Results & Metrics
Significantly boosted Mean Reciprocal Rank (MRR) and precision over the baseline model
Achieved robust, low-latency API response times suitable for real-time production
Streamlined resume parsing and candidate matching into a single cohesive pipeline
Lessons Learned
Hybrid search (semantic + sparse) is vital for domain-specific jargon like skillsets
Multi-agent workflows require strict prompt engineering and validation constraints
In-memory caching is essential for deploying ML-heavy backends at scale