Vetted Machine Learning Engineers in the Greater Boston

“LLM-focused customer-facing engineer who repeatedly takes document Q&A and agentic prototypes into secure, monitored production systems. Experienced in reducing hallucinations via RAG + guardrails, diagnosing retrieval/embedding issues in real time, and partnering with sales to run metrics-driven PoCs that overcome accuracy/security objections and drive adoption.”

“Built and shipped a real-time oncology risk prediction system used by doctors during patient visits, trained on clinical data in AWS SageMaker and deployed via FastAPI with sub-second responses. Emphasizes clinician-trust features (SHAP explainability, validation checks) and HIPAA-compliant controls (encryption, RBAC, audit logging), plus Kubernetes-based production operations with autoscaling, monitoring, and drift/retraining workflows; collaborated closely with oncologists at Flatiron Health.”

“Built and deployed a fine-tuned Qwen 2.5 14B model into production at Dextr.ai as the backbone for hotel-operations agentic workflows, running on AWS EKS with Triton and TensorRT-LLM. Demonstrates strong cost-aware LLM engineering (QLoRA, FP8/BF16 on H100) plus rigorous benchmarking/observability (Prometheus, LangSmith) with reported sub-30ms TTNT. Previously handled long-running ETL orchestration with Airflow at GE Healthcare and Lowe's.”

“AI engineer who deployed a production LLM-powered safety system for an education platform, combining rule-based checks, multi-LLM verification, and selective context (prompt+image vs image-only) to prevent explicit prompts/images from getting through. Strong focus on reliability via benchmarking, trace-based failure analysis, and continuous improvement driven by stakeholder feedback and manual review.”

“Built and deployed a production LLM-powered biomedical knowledge extraction pipeline that processed millions of papers to identify tools/techniques and produce a unified knowledge graph via active learning NER (Prodigy + spaCy transformers) and entity linking (Bio-tools/Wikidata). Addressed hard NLP engineering challenges like WordPiece span-offset alignment and scaled inference over ~1.5M documents using batching/caching, containerized services, async workers, and orchestration with Prefect/Airflow.”