COLLINS WESTNEDGE

SKILLS

Programming: Python (PyTorch, Transformers, NumPy, pandas, scikit-learn), SQL, Git

LLM/NLP: RAG, prompt engineering, fine-tuning (mixed precision, grad accumulation), shallow fusion for ASR, activation steering, debiasing, contrastive learning

MLOps & Data: Experiment design & permutation tests, calibration, distributed training, MLflow, Databricks/Spark, Azure (Cognitive Search, Document Intelligence), Kubernetes, Hugging Face Hub

EXPERIENCE

• Led data-science efforts for a CSR-facing RAG platform; designed & implemented automated evaluation reducing A/B cycles from ~30 days to 7 minutes; adopted across CSR org (~700 users)
• Unlocked enterprise open source access by partnering with risk/security and documenting guardrails; delivered Whisper ASR pipeline with 95% cost reduction vs AWS Transcribe and 30% relative WER improvement over vendor
• Engineered a multi-stage LLM workflow for regulatory checks (risk tolerance & consent) with human-in-the-loop validation; saved 200+ hours annually while meeting audit requirements
• Sponsored & advised InfoNCE fine-tuning (contrastive + domain adaptation); project owner achieved a 70% reduction in vector storage

• Reviewed & advised ML/AI grants, guided experimental design and evaluation, and delivered workshops and panel discussions on ML/AI

• Analyzed commercial real-estate datasets to inform portfolio allocation and risk modeling for institutional clients

SELECTED PROJECTS

• Implemented shallow fusion integrating Whisper ASR with domain-adapted GPT-2 via a custom LogitsProcessor; achieved an 8.5% relative WER reduction (p = 0.012) on medical speech using beam-search integration
• Built an evaluation framework with medical terminology normalization, permutation-based significance testing, and domain-specific error analysis; released models and complete implementation

• Fine-tuned GPT-2 on 3.6B PubMed tokens using a production-grade training pipeline with mixed precision (FP16), gradient accumulation (128 effective batch size), and fault-tolerant checkpointing
• Implemented training optimizations including cosine scheduling with linear warmup, selective weight decay capability, and automatic checkpoint recovery; achieved stable convergence at 131K tokens/batch

• Developed a debiasing method using projection-based concept removal in embedding spaces; 42% reduction in bias triggers while maintaining comparable retrieval quality

EDUCATION