Professional Overview
I built VSCE-EL, a five-stage multimodal entity linking pipeline over 90K+ enti- ties. It combines hybrid retrieval, cross-encoder reranking, and LoRA-adapted LLM selection. I fine-tuned Qwen2-7B via QLoRA (5.0M trainable parameters, 0.066% of 7.6B) with curriculum training and diagnosed a position-bias failure in the reranker, which I corrected through listwise softmax fine-tuning. The system reaches Top-1 accuracy of 94.75% on WikiMEL, 93.38% on Richpedia, and 93.12% on Wikidiverse. Two other projects shaped how I think about production systems. DataPilot is a six-agent LangGraph pipeline coordinated through three MCP tool servers, with a retry loop that recovers from SQL errors automatically. I built server-side read-only SQL enforcement and ChromaDB business-term grounding into it, which blocked every adversarial SQL-injection attempt I tested regardless of how the LLM behaved, and an eval harness that surfaced hallucination-class errors I would have missed reviewing outputs by hand. WebRAG is a local-first RAG system using dense and lexical retrieval merged via Reciprocal Rank Fusion, with local LLM reranking and no cloud API dependency, self-hostable through Docker Compose.
Areas of Expertise
AI & Machine Learning Engineers
Design, train, and ship production ML models — from data pipelines to model inference at scale. Covers supervised learning, neural networks, LLM integration, and model optimisation.
AI Product & Strategy
Translate AI capabilities into product roadmaps, define success metrics, and drive adoption across engineering and business teams.
MLOps & AI Infrastructure
Architect scalable ML infrastructure — CI/CD for ML, model versioning, experiment tracking, monitoring, and cloud deployment on AWS, GCP, or Azure.
