Muqarrab Rahman

Summary

What I did

• Designed the scoring stack from AI signals across CV-to-JD skill matching, job-responsibility alignment, and functional-experience depth, emitting confidence values and reasoning traces at each stage while maintaining PII-compliant processing.
• Extended the same matching engine to reverse AI job sourcing and adapted the architecture for the Department of Government Enablement (DGE) and Remote Work Emirates Foundation use cases.
• Built bilingual English/Arabic resume-to-JD scoring across all four language permutations, combining duration-weighted functional experience, skill matching, and career-level distance; generalized seniority into a six-tier taxonomy and mapped enterprise skill banks to ESCO.
• Standardized retrieval using BGE bi-encoders, cross-encoder reranking, and FAISS; reduced LLM spend in the CV-parsing pipeline by moving structured extraction to smaller models with JSON tool calls, reusing parsed fields downstream, and adding prompt/tool caching.
• Architected reusable multi-agent workflows using LangGraph and LangChain with tool calling, state management, memory, conditional routing, context retention, inter-agent communication, and multi-step orchestration for enterprise document-processing and conversational-AI use cases.
• Built an LLM-based interview agent that conducts structured, multi-turn candidate interviews, generates contextual follow-up questions, and evaluates responses in real time.
• Developed a workspace-connected AI agent integrated with internal APIs, enabling users to query dashboard and workspace data in natural language without manual report lookup.
• Introduced a staged orchestration approach using deterministic/rule-based filtering and candidate scoring for initial screening to reserve LLM calls for steps requiring deeper reasoning.
• Designed a hierarchical agent architecture using LangGraph where a supervisor/router interpreted user intent to identify relevant HCM lifecycle modules.
• Implemented a shared state as a control plane between agents to carry conversation context, identified entities, and results from downstream agents.
• Managed cross-module queries by routing state sequentially through specialized agents, such as recruitment and interview-scheduling agents.
• Used explicit state transitions and checkpoints with structured outputs to ensure predictable handoffs and allow for retries or requests for missing information.

Results

• Ranked a corpus of 20,000+ active applicants for Mubadala Investment Company.
• Reduced LLM spend in the CV-parsing pipeline by moving structured extraction to smaller models with JSON tool calls and prompt/tool caching.

Coursework