Vetted Hugging Face Professionals

“At PwC, built and productionized an agentic RAG enterprise search assistant over 6M internal documents (8M embeddings), deployed across AWS and GCP. Drove major retrieval gains (72%→92% precision via BM25+dense hybrid with RRF and cross-encoder re-ranking), reduced hallucinations 30%, achieved <2s latency at 50–60K queries/month, and cut support tickets 30%—boosting adoption to 2,500 users by adding source-cited answers.”

“Robotics ML/software engineer focused on Vision-Language-Action control for 7-DoF robots, replacing tokenized action decoding with continuous regression heads (including a logit-weighted expectation approach) to improve stability and real-time behavior. Strong in ROS1/ROS2 systems integration and debugging closed-loop manipulation issues via latency instrumentation, QoS-aware distributed messaging, and sim-to-real validation using Gazebo/Unity, Docker, and CI pipelines.”

“At Esri, the candidate is building a production LLM-powered WebGIS AI framework that embeds an AI assistant into web maps and routes natural-language requests into ArcGIS JavaScript SDK functions via a LangGraph-orchestrated, multi-agent system. They emphasize production reliability and scale (strict tool calling/JSON, live schema validation, query guardrails) and rigorous evaluation/observability using LangSmith, offline prompt datasets, and latency/tool-call accuracy tracking.”

“Built and deployed a production RAG-based LLM Q&A and summarization platform for internal documents, emphasizing grounded answers with structured prompting and citations to reduce hallucinations. Experienced orchestrating end-to-end LLM workflows with LangChain plus cloud pipelines (Azure ML Pipelines, AWS), and runs iterative evaluation using both metrics (accuracy/hallucination/latency/cost) and real user feedback to drive reliability.”

“ML/LLM practitioner with experience at Truveta building an LLM-based evaluation framework; identified non-overlapping evaluator failure modes and proposed an ensemble approach that enabled scaling training data and drove ~5% performance gains across multiple internal projects. Strong focus on robustness to distribution shift (augmentation/domain adaptation/meta-learning) and production reliability via monitoring, drift detection, and safe fallbacks.”

“Built and deployed a live LLM-powered platform that takes a LinkedIn job URL + resume and generates job-specific resumes and personalized outreach at scale, with production-grade logging/monitoring/retries on Vercel + Railway. Experienced with agent orchestration (AWS Bedrock/Strands, LangGraph, CrewAI) and rigorous AI workflow testing, plus stakeholder-facing prototypes like data lineage/metadata and NL-to-SQL + dashboard generation.”

“Industrial AI/ML practitioner with experience deploying real-time monitoring and anomaly detection in a regulated Sanofi vaccine manufacturing facility, including root-cause workflows, logging/alerting, and SOP-aligned validation—achieving ~90% faster anomaly detection. Also built Python/NLP-style automation to accelerate instrumentation & control documentation (~40% faster) and delivered end-to-end predictive analytics for an agri-food operations/distribution client using close operator and leadership feedback loops.”

“Backend/AI engineer at Intuit who built a production AI-powered case assistant for support agents (FastAPI on AWS EKS) combining Postgres case data, OpenSearch retrieval with embedding reranking, and internal LLMs. Improved peak-season reliability by diagnosing P95/P99 timeout spikes and cutting P95 latency from ~800ms to <400ms via composite indexing, keyset pagination, connection pool tuning, and caching, while adding grounded-generation guardrails (evidence packs, confidence thresholds, fallbacks, human-in-the-loop).”

“Applied LLMs and a graph-RAG architecture in Neo4j to automate an accounting firm's cross-checking of transactional books against tax regulations, indexing 1,000+ pages into a knowledge graph with vector search. Combines agentic LLM workflows with classical NER (Hugging Face/NLTK) and validates using expert-labeled held-out data plus precision/recall and measured accountant time savings after deployment.”