Shubham Dubey

Production LLM Stack — Healthcare Intelligence Platform

• Architected a GPU-clustered LLM inference stack: 5 production inference pipelines, 2 continuous fine-tuning loops, and 2 automated eval gates serving millions of patient records.
• Taxonomy Classification: 3-stage Gemma-4-E4B-it pipeline with FAISS-GPU + EmbeddingGemma-300M MMR vector search (λ=0.9), structured breed/species classification with per-match QA scoring. Supports multi-breed resolution (e.g., “Labrador Poodle mix” → 2 matches) with intelligent retry logic (soft/hard caps).
• Gender Normalization: Deployed Gemma-4-E4B-it via vLLM to standardize raw descriptions into 7 clinical categories with plausibility scoring, including domain rules (gelding → Male Neutered horse, barrow → Male Neutered pig).
• Transcription & SOAP QA: Operationalized Gemma-4-31B-IT-QAT-W4A16 via Ray + vLLM for clinical transcript evaluation (quality, plausibility, completeness) and SOAP summary faithfulness scoring with hallucination/omission/medication error detection.
• MLOps & Eval Gates: Closed-loop fine-tuning where candidate models are evaluated on holdout accuracy against production; promoted via MLflow alias only on statistically significant improvement. Orchestrated via dbt Python models on L40S GPU clusters with Unity Catalog governance.

Enterprise Data Lakehouse & Architecture

• Architected the enterprise data lakehouse on AWS (S3, Lambda, EMR, EKS) using Databricks and Snowflake, ingesting 50TB+ weekly from CRM, ERP, PIMS, and HR systems into columnar and relational stores.
• Pioneered Data Contracts and schema evolution protocols with automated quality, anomaly detection, and lineage tracking ensuring 100% integrity for executive reporting.
• Designed SCD Type 2 semantic models in dbt/Snowflake managing 15+ years of historical KPIs; conducted Level-of-Detail (LOD) analysis in Python to isolate revenue drivers and longitudinal trends.
• Built star schemas and semantic models reducing transformation runtimes by 40% and accelerating financial/operational reporting for enterprise stakeholders.

Event-Driven Streaming & Real-Time Analytics

• Designed config-based streaming infrastructure over Apache Kafka and Spring Cloud Data Flow, processing millions of events per second with sub-second latency. Eliminated point-to-point connections, cutting infrastructure costs by 5%.
• Architected an end-to-end streaming control plane using Kafka, Redis, Snowflake, and Python/Java microservices for real-time predictive scoring achieving 92% accuracy; integrated automated health alerting via Splunk/CloudWatch.

Integrations, Security & Observability

• Built resilient CDC connectors for Salesforce and NetSuite via OAuth 2.0 with least-privilege IAM governance, improving real-time data accuracy by 25%.
• Established enterprise observability with CloudWatch, Splunk, and custom validation frameworks, reducing data drift and service errors by 50% and maintaining 99.9%+ availability.

Platform Engineering & DevOps

• Automated cloud provisioning via Terraform, Docker, Kubernetes, and GitHub Actions, establishing GitOps CI/CD that reduced deployment times by 60%.
• Standardized microservice containerization enabling rapid, iterative AI feature deployment across distributed teams.

• Re-engineered legacy ETL into parallel asynchronous ELT pipelines in Python/Java, accelerating query speeds and Tableau refresh rates by 45% for global supply chain and finance stakeholders.
• Designed Type 2 SCD dimensional schemas in Snowflake tracking longitudinal workforce metrics, trial outcomes, and historical performance with complete audit trails.
• Implemented automated validation and anomaly detection engines, eliminating 50% of data quality defects prior to downstream ML model and executive report consumption.