Senior Machine Learning Engineer

I developed an agentic RAG platform for call analytics that answers questions about customer refusals, outcomes, technical issues, objections, and closed deals. I built a LangGraph planner, SQL tooling for reliable counts, and citation-backed synthesis. The retrieval stack combined HyDE, lexical expansion, e5, BM25, reciprocal rank fusion, and bge reranking, improving Recall@20 from 0.61 to 0.83 and reducing analytical report preparation from days to minutes.

I built a large-scale speech analytics pipeline for approximately 40,000 contact-center calls per day and 650 monthly hours of video meetings. The platform combined fine-tuned GigaAM ASR, pyannote diarization, TF-GridNet speech separation, and Qwen-based classification. It raised quality-control coverage from 0.3% to 100%, while violation detectors achieved macro F1 of 0.83.

I delivered an agentic document workflow system that extracts, normalizes, validates, and maps entities into report forms with source traceability. The solution reduced form completion time from approximately 25 minutes to 3 minutes and achieved 96% field-level accuracy.

I also built a regulatory RAG system over approximately 180,000 banking and regulatory documents for around 2,500 employees. The system used OCR, e5 embeddings, FAISS, Elasticsearch BM25, RRF, bge reranking, and Llama 2 generation with verified citations, reducing response time from 25 minutes to 1.5 minutes.

I developed computer vision solutions for automotive production lines, including instance segmentation models to detect vehicle parts and mounting points and verify correct installation. I used class-balanced sampling and image augmentation for production conditions, exported models to ONNX for edge inference, and achieved mIoU of 0.89 with 0.94 precision for installation-defect detection.

I implemented multi-object tracking and production traceability across manufacturing stations, linking operations to individual vehicles and time windows. I also built OCR and NER pipelines for VINs, labels, and work orders, producing complete vehicle reports and reducing report preparation from around 30 minutes to under one minute. I owned the full ML lifecycle from problem formulation and data collection through deployment of high-load models on Azure.

I architected and implemented an MVP video analytics service for streaming data recording and real-time event classification. I worked on feature extraction methods using vector analysis and physical functions, including divergence and vector flows, to identify semantic transitions in data.

I improved temporal action localization performance by optimizing semantic vectorization, increasing mAP by 5 percentage points. I presented the mathematical foundations of this work at IEEE ISIE 2021 in Kyoto.