• Constructed real-time product-ranking models using PyTorch, feature stores, and Snowflake signals, improving conversion

lift by 6.8% across high-traffic categories through optimized inference pipelines and continuous monitoring workflows.

• Architected demand-forecasting system integrating PySpark pipelines, incremental Delta ingestion, and SHAP analysis,

reducing weekly inventory volatility by 12% and stabilizing replenishment across electronics and seasonal assortments.

• Engineered anomaly-detection framework combining embeddings, Kafka inputs, and drift-detection metrics, detecting

pricing irregularities 23% faster and safeguarding pricing experimentation across merchandising clusters.

• Developed retrieval-grounded assistant using vector embeddings, RAG routing, and structured prompts, decreasing agent

resolution time by 18% for warranty, billing, and device-troubleshooting queries.

• Refined multi-agent LLM orchestrations for search-intent rewriting, improving relevance ranking by 9.4% across long-tail

queries through prompt specialization, model pruning, and contextual grounding from historical navigation logs.

• Implemented response -safety filtering with quantized instruction-tuned models, reducing hallucination rates 31% and

ensuring compliant customer assistance workflows across enterprise geographies and product categories.

• Designed an item-embedding pipeline using Scala, PySpark, and Delta Lake to cluster SKUs by behavioral affinity,

improving cross-category recommendation depth by 14% during seasonal campaigns.

• Built scalable candidate-generation models with hierarchical sampling and ANN retrieval, raising same-session discovery

rates by 9% while decreasing tail-exposure gaps across diverse merchandise hierarchies.

• Constructed a feature store-backed training fabric integrating transactional events, feed signals, and contextual

metadata, shortening model refresh latency from 22 hours to 9 hours during peak traffic.

• Deployed sequence-based re-ranking models in Triton inference servers with optimized batching heuristics, reducing P95

latency by 31% and sustaining stable throughput during mega-sale concurrency surges.

• Developed synthetic-cart augmentation using behavioral perturbations and constrained sampling, enhancing robustness

to sparse-item interactions and improving add-to-cart prediction recall by 6% across underrepresented assortments.

• Implemented continuous-monitoring dashboards with drift metrics, embedding-health diagnostics, and slice-based

stability alerts, enabling rapid anomaly triage and maintaining consistent performance in dynamic retail patterns.

• Collaborated with merchandising analysts to validate model shifts through uplift experiments, aligning algorithmic

behaviors with commercial objectives and ensuring compliant rollout of recommendation updates across categories.