Utkarsh Gupta

LLM Agents • Multi-source RAG • Dynamic Skill Orchestration • AI Governance

Agentic fraud investigation platform combining LLM reasoning, multi-source retrieval, dynamic skill dispatch, and automated governance.

• Two-phase workflow in LangGraph—stateful nodes and conditional edges keep escalation auditable and replayable; low-risk events gated deterministically, ambiguous cases escalated to the LLM branch, controlling cost and latency.

• Registry-driven skill architecture enabling dynamic selection and parallel fan-out across Presto, ClickHouse, and S3, with reducer-based state merge for concurrent outputs.

• Grounded Amazon Bedrock/Kimi K2 reasoning in real-time evidence, enforcing structured JSON contracts, score calibration, credential isolation, and audit-ready traces.

• Spring Boot microservice fronting the orchestration layer - owns REST contract, auth, credential brokering, and auditing, decoupling the agent layer from the enterprise layer.

• Evaluation harness: 90% investigation accuracy vs. 78–83% for isolated skills; orchestration resolved 71.4% of conflicting skill assessments.

• Token, latency, execution, and retrieval-level observability for cost governance and production monitoring.

Production ML • CatBoost • SageMaker • Drift Detection • Automated Retraining

Led the ML/MLOps lifecycle for an account-takeover fraud detection platform - feature engineering, training, real-time inference, monitoring, and retraining.

• Built an end-to-end CatBoost pipeline on SageMaker covering feature engineering, training, batch inference, and containerized real-time inference.
Led the retraining pipeline across MWAA/Airflow, AWS Glue, SageMaker, and ClickHouse.

• Designed automated model/data monitoring using PSI, Kolmogorov–Smirnov, and χ² tests, with Datadog alerting, Evidently reports, Streamlit dashboards, and a 38-test validation suite.

• Maintained production MWAA workflows for fraud metrics, ATO detection, and Kafka event publishing, with OpenTelemetry observability.

• Used AWS MCP servers with Claude to accelerate deployment and ops — inspecting SageMaker endpoints, Glue jobs, and MWAA DAG state through a schema-defined tool interface instead of ad-hoc scripting.

• Exposed ClickHouse fraud metrics via MCP for conversational drift and performance queries during incident triage, with scoped read-only credentials keeping access auditable.

Computer Vision • AutoML • Defect Classification • Production Deployment

• Developed production ML and computer-vision solutions for nanometer-scale semiconductor defect inspection, automating defect filtering, classification, and attribute extraction across high-volume wafer inspection data.

• Built an AutoML/scikit-learn defect classification pipeline, improving defect filtering performance by 35%.

• Engineered image-processing pipelines for automated defect attribute extraction, achieving around 95% faster processing per tile.

• Developed segmentation modules for patterned semiconductor images to improve region-of-interest identification and downstream defect analysis.

• Deployed containerized ML pipelines using Docker, enabling plug-and-play training and inference in customer/on-site environments.

• Implemented ONNX-based production inference graphs optimized for high-throughput execution.

RAG • LLMs • Conversational AI • Knowledge Retrieval • Prompt Engineering

Designed and developed production-oriented RAG and conversational AI systems combining knowledge retrieval, LLM generation, prompt engineering, and multi-step conversational workflows.

- Built the end-to-end RAG lifecycle covering data ingestion, extraction, chunking, normalization, embeddings, retrieval, context construction, and LLM generation.

- Designed conversational state-management strategies to preserve relevant context across multi-turn interactions while minimizing context pollution and hallucination.

- Implemented retrieval and prompt optimization strategies to improve grounding, response quality, and knowledge utilization.

- Integrated knowledge retrieval with conversational workflows to select the appropriate knowledge sources and tools based on interaction state.

- Applied production-oriented evaluation and monitoring approaches to assess retrieval quality, hallucination, and response reliability.

90B Rows/Month • AWS Data Engineering • ML Infrastructure • Cost Optimization

• Designed and optimized a cloud-based data platform processing approximately 90 billion rows per month to provide scalable, reusable data products for analytics and ML workloads.

• Built large-scale ETL pipelines for ingestion, normalization, metric computation, and downstream ML/analytics consumption.

• Optimized AWS Glue/Spark processing to reduce DPU costs by ~40–50%.
Improved data-read performance by approximately 5× and reduced query execution time by ~70% through storage and data-model optimization.

• Worked across AWS Glue, S3, Snowflake, Presto/Athena, Redshift, and ClickHouse.

• Designed reusable data pipelines supporting a single-source-of-truth architecture for business and technical metrics.

Active Learning • Human-in-the-loop ML • Data Efficiency

• Designed an active-learning workflow to reduce the manual labeling burden associated with hundreds of thousands of semiconductor inspection defects.

• Developed an intelligent sample-selection strategy to prioritize the most informative defects for expert labeling instead of relying on random sampling.

• Reduced the effective labeling workload from approximately 500,000 candidate defects to around 500 targeted samples.

• Reduced mislabeled samples by approximately 40% while improving system accuracy by around 25%.

• Created an iterative feedback loop between model predictions and expert annotations to continuously improve defect classification quality.