Samuel Kh

Designed a Graph RAG system pairing Neo4j graph traversal with Pinecone embedding search, improving answer completeness by 28% over a vector-only baseline on a 200-question test set. Fine-tuned open-source LLMs with QLoRA and DPO transformers for structured extraction, improving output consistency from 76% to 94% on a 600-sample test set in one quarter, with reduced effort via LangSmith scoring. Built a LangGraph multi-agent workflow orchestrated with OpenAI/Azure OpenAI and Cloud multi-step research, reducing an 8-minute task to under 3. Created a voice AI service using Whisper and FastAPI for transcription, classification, and summarization across 12,000 conversation minutes monthly, reducing human review by 60%. Deployed ML/GenAI services as REST APIs on Docker, Kubernetes, and SageMaker with OpenTelemetry, and led a team by running model reviews, enforcing GDPR/CPA practices, and mentoring engineers across agile sprints.

Developed HIPAA-compliant pipelines converting prescriptions, medical bills, and clinic notes into usable fields, processing 18,000 documents monthly with 93% accuracy while reducing hand entry by 65% in 16 weeks. Engineered real-time anomaly detection on Kafka and PySpark with scikit-learn, scoring 1,500 events per second and lowering false positives by 24% over 5 months with Power BI alert trend tracking. Trained XGBoost classifiers on SQL tables, increasing precision by 18% using feature engineering, hyperparameter tuning, and error analysis. Scaled distributed training and input processing on Databricks with Delta Lake across AWS, Azure, and GCP, reducing training from 7 hours to 2.5. Automated an Airflow ETL validation/training/experiment tracking workflow, trimming repetitive ML operations by 50% in 16 weeks, and built speech-to-text and NLP for patient calls covering 8,000 audio minutes per month with 70% reduction in transcription effort.

Developed SQL, scikit-learn, XGBoost, and Keras models for classification, regression, and forecasting for corporate clients, improving baseline performance by 17% and reducing false positives by 15% in the first half-year. Built Spark preparation jobs handling 2 million records daily, shrinking preparation time by 45% and speeding experimentation. Applied feature selection, cross-validation, hyperparameter tuning, and error analysis with Pandas and NumPy, reducing iteration time from 5 days to 3. Created FastAPI and Docker inference APIs that shortened integration from one week to three days, and moved experimental ML work into repeatable production runs, removing 40% of manual steps.