Synapse
LLM-guided genetic resume optimization for explainable job-person fit
Synapse is a two-phase retrieval and resume optimization system that improves job-person fit through explainable ranking and LLM-guided genetic evolution of resumes toward target job postings.
How it works:
- Phase 1 retrieves candidate job postings using semantic and keyword-based signals
- Phase 2 applies LLM-guided Differential Evolution to iteratively improve resume alignment with target postings, using mutation, crossover, and fitness evaluation operators
- The system surfaces fine-grained skill and keyword alignment explanations at each step
Key results:
- Weighted-average rank ensemble achieves nDCG@10 of 0.714, a +31.9% improvement over the embedding-only baseline (0.541)
- Contrastive learning reranker alone yields +17.4% nDCG@10; LLM pairwise ranking adds +10.2%
- GPU (H100) accelerates Phase II scoring 22× over CPU (0.026s vs. 0.570s per query)
- LLM-guided evolutionary resume optimization achieves a median 62%, mean 68%, upper-quartile 92% fitness gain over 5 generations across 100 resumes
- Corpus: 120,000 LinkedIn job postings; 2,500 LiveCareer candidate resumes
Stack: Python, LLMs (API), vector retrieval, evolutionary algorithms
Publication: Erol, A. K., Yoon, S., Hom, K., & Zhang, X. (2026). Synapse: Evolving Job-Person Fit with Explainable Two-phase Retrieval and LLM-guided Genetic Resume Optimization. arXiv preprint.