Enterprise AI Programs

Business Problem

Enabling American Express to operationalize Generative AI at enterprise scale — reliably, securely, and with governance built in — transforming how the organization uses AI across products, operations, and customer experience.

Architecture

• Enterprise GPU AI Platform for LLM inference and training
• RAG pipelines with enterprise knowledge grounding
• Agentic AI systems for multi-step business workflows
• AI Governance layer: safety, audit, compliance monitoring
• LLMOps: evaluation, versioning, and observability

My Role

• Led design and delivery of enterprise AI platform components
• Defined architecture for production GenAI and Agentic systems
• Partnered with product and engineering on AI roadmap
• Drove responsible AI and governance frameworks

AI Techniques

• GPT-5 and frontier LLM deployment
• Retrieval-Augmented Generation (RAG)
• Agentic AI with tool use and planning
• Prompt engineering and evaluation
• Fine-tuning and RLHF

Business Impact

• Enterprise-scale AI capability enabling multiple business units
• Accelerated GenAI adoption across the organization
• AI governance framework ensuring responsible deployment
• Measurable improvements in AI-powered customer experiences

Key Lessons

Building Enterprise GenAI is fundamentally a platform and governance challenge, not just a model challenge. RAG quality, evaluation rigor, and agentic safety are the hard problems.

Business Problem

A major health insurance provider needed an AI-enabled physician consultation platform — integrating diagnostics, telehealth, patient monitoring, and NLP — to improve healthcare access and physician decision-making for 40M+ members.

Architecture

• HealthOS: integrated AI platform for consultation and monitoring
• Computer vision pipeline: X-ray, CT, pathology analysis
• Live video analytics for patient anxiety and expression
• NLP chatbot for FAQ automation and patient support
• Real-time vitals alerting for post-operative patients
• AWS Lambda, Glue, SageMaker cloud infrastructure

My Role

• Led the AI Tower — full delivery of the AI platform
• Defined architecture across image, text, and streaming AI
• Managed cross-functional teams across 3 AI workstreams
• Drove production launch and clinical validation

AI Techniques

• Disease detection from X-ray and CT images (TensorFlow, CNN)
• Pathology and vitals analytics
• NLP chatbot for healthcare FAQ (BERT-based)
• Live video anxiety detection (OpenCV, deep learning)
• Nail-image health screening (computer vision)

Business Impact

• Supported 40M+ healthcare members with AI-augmented care
• Reduced physician consultation time with automated diagnostics
• Enabled remote post-operative monitoring at scale
• NLP chatbot resolved 60%+ of member queries without escalation

Key Lessons

In healthcare AI, clinical trust is the hardest problem. Explainability, validation, and strong human-in-the-loop design are non-negotiable — not optional features.

Business Problem

Thyssenkrupp needed to predict elevator failures before they happened across 150K+ units globally — reducing unscheduled downtime, lowering maintenance costs, and improving customer service levels.

Architecture

• Azure Data Lake + Databricks platform for IoT ingestion
• Daily elevator-level feature engineering at massive scale
• 7-day ahead failure prediction models per elevator unit
• Maintenance alert system integrated with TKE web services
• Tableau and Power BI dashboards for operations teams
• Azure CI/CD for automated model deployment

My Role

• Led data science delivery end-to-end
• Designed feature engineering framework for 150K+ elevators
• Built and validated supervised failure prediction models
• Scaled platform from 10K to 200K units

AI Techniques

• Supervised learning (XGBoost, Random Forest, gradient boosting)
• Time-series feature engineering at IoT scale
• Class imbalance handling for rare failure events
• PySpark for distributed model training
• Model monitoring and drift detection in production

Business Impact

• Platform scaled to 200K+ elevators across 3 countries
• Significant reduction in unscheduled maintenance events
• Maintenance teams alerted 7 days before predicted failures
• Direct improvement in customer uptime SLAs

Key Lessons

At IoT scale, the challenge is not the model — it is the data pipeline, feature reliability, and operational integration. Production AI is an engineering problem first.

Business Problem

Pratt & Whitney needed to predict aeroengine combustion distress events before they occurred — enabling proactive maintenance, improving engine reliability, and supporting engineering decision-making for safety-critical systems.

Architecture

• Sensor data ingestion and time-series processing pipeline
• Combustion distress forecasting models
• Real-time diagnostic inference engine
• Text extraction from asset documents and scanned images
• Engineering decision support dashboards

My Role

• Designed and built prognostic forecasting models
• Built real-time diagnostic systems for engine events
• Developed document and image text extraction tools
• Worked directly with aerospace engineering teams

AI Techniques

• Time-series anomaly detection and forecasting
• Recurrent neural networks for sensor sequence modeling
• Image-based text extraction (OCR, deep learning)
• Signal processing for sensor noise reduction

Business Impact

• Early detection of combustion distress improved maintenance planning
• Diagnostic capability reduced engineering investigation time
• Document intelligence automated asset information capture

Key Lessons

In safety-critical systems, model uncertainty quantification and human-in-the-loop validation are not optional. Trust must be earned through rigorous validation before operational deployment.

Business Problem

Transportation, utilities, and infrastructure clients needed automated intelligence from satellite imagery, mobile mapping data, and 3D LiDAR scans — replacing expensive manual workflows with AI-powered asset detection and classification.

Architecture

• Scalable CV pipeline: satellite, mobile mapping, LiDAR
• Road marking and building footprint extraction
• Land-cover and land-use classification system
• POI detection and privacy-preserving anonymization
• 3D LiDAR processing with PointNet/Open3D
• Azure Databricks for distributed processing

My Role

• Led delivery of CV platform across multiple client verticals
• Built and managed AI team of 30+ engineers and scientists
• Defined CV architecture strategy across imaging modalities
• Drove client engagement and solution design

AI Techniques

• Semantic segmentation (U-Net, DeepLab)
• Object detection (YOLO, Faster R-CNN)
• 3D point cloud processing (PointNet, Open3D)
• GAN-based data augmentation for training
• Multi-resolution imagery fusion

Business Impact

• Automated asset mapping at 20–40× the speed of manual workflows
• Deployed across automotive, utilities, telecom, aerospace clients
• LiDAR vegetation detection enabled proactive grid maintenance
• Privacy-preserving blurring met GDPR requirements for EU clients

Key Lessons

Multi-modal CV at geospatial scale requires strong data pipelines, careful annotation quality control, and domain-specific validation — not just model architecture innovation.