Hi, I am Raj

I work on trade-surveillance AI for equity markets, where millions of market events must be monitored in near real time to detect suspicious behavior.

My role focuses on building explainable, production-grade detection systems that improve recall while satisfying regulatory model-risk standards.

• Built graph-based anomaly detection workflows on ~3.2M daily trade relationships, engineering structural and temporal features that improved recall on regulatory-flagged scenarios by 34%.
• Developed sub-second streaming detection modules on order-book feeds (~200K events/sec), combining event aggregation and low-latency model scoring for faster monitoring of abnormal trading behavior.
• Implemented explainability pipelines using SHAP and attention-based attribution to generate audit-ready justifications and support SR 11-7 model-risk review processes.
• Partnered with quant research and compliance to incorporate alternative signals (news sentiment and options flow) into feature-store pipelines, improving event-driven manipulation detection by 27% in backtesting.

I built AI systems for clinical-trial matching and performance marketing, where ranking quality and conversion efficiency directly impacted customer outcomes.

I owned model development, experimentation, and monitoring workflows across multiple production deployments with measurable business targets.

• Built an end-to-end patient-trial ranking pipeline using Sentence-BERT embeddings and LightGBM, improving matching precision by 50% across 10+ deployments.
• Designed and analyzed A/B experiments for ranking and ad-generation strategies, using funnel metrics and statistical testing to double qualified conversions while reducing campaign spend by 30%.
• Created monitoring dashboards for CTR, CPL, fairness, drift, and funnel performance, enabling rapid diagnosis of model and campaign changes for engineering and business teams.
• Supported KPI tracking and client-facing reporting across 6+ implementations, contributing to roughly $500K in revenue through data-driven optimization.

I developed a scalable data and AI integration ecosystem to streamline how climate-education resources were ingested, validated, and published.

My role combined NLP modeling, API engineering, ETL automation, and experimentation to improve both operational efficiency and user outcomes.

• Architected an NLP-driven automated classification pipeline using Python and LLM-based models, reducing manual curation effort by 70% across 1,000+ resources through automated tagging.
• Engineered backend APIs and end-to-end ETL workflows with FastAPI and PostgreSQL, supporting the processing and publication of 100+ weekly opportunities with automated schema enforcement.
• Implemented data validation checkpoints and reliability controls across ingestion pipelines, improving data consistency and reducing downstream publishing errors.
• Used behavioral analytics and A/B experimentation to refine content strategies, contributing to a measured 25% increase in user satisfaction metrics.

I developed NLP and anomaly-detection systems for enterprise IT operations, where prediction quality directly affected routing speed, SLA risk, and support efficiency.

I focused on scalable modeling, feature engineering, and reliable MLOps practices for high-volume production workflows.

• Built a multi-label ticket-classification system on 2.3M ServiceNow records using fine-tuned BERT, achieving 91.4% micro-F1 and reducing MTTA by 68% (~14,000 hours saved annually).
• Developed escalation-risk predictors with XGBoost and engineered entity, text, log, and temporal features, reaching 87% accuracy for proactive incident intervention.
• Implemented streaming anomaly detection with Isolation Forest for SLA-risk identification, achieving 94% recall across real-time operational workflows.
• Designed validation and monitoring workflows for production models, reducing model-related incidents by 43% through tighter quality controls and performance tracking.

I engineered a physics-informed machine learning framework to automate gap-filling of greenhouse gas (CO₂) flux data for climate research workflows.

This work is in publication stage and focuses on producing auditable, high-quality imputed signals usable by global environmental researchers.

• Engineered 14+ high-dimensional feature groups, including multi-scale lags, STL decomposition, and physics-based transforms (Q10 temperature response and moisture sigmoid), improving non-linear signal detection by 25% versus raw-data baselines.
• Architected a multi-regime cascade pipeline with LightGBM and Random Forest that adapts feature selection to information decay, sustaining predictive quality across missing gaps from hours to 70+ days.
• Built a robust imputation validation framework with ablation studies and missingness indicators so models explicitly distinguish observed vs. imputed states for auditability.
• Delivered a publication-grade ML workflow that automates CO₂ flux reconstruction and supports reproducible downstream climate-analysis pipelines.

I architected a high-velocity computer vision pipeline to automate vehicle tracking and license plate recognition in real time under challenging conditions.

I focused on model-ensemble design, inference optimization, and robust feature engineering to achieve production-ready throughput and accuracy.

• Engineered a multi-stage ensemble pipeline using YOLOv8/YOLOv5, DeepSORT, and CRNN, achieving 0.97 F1-score and 0.99 recall for automated vehicle identification and tracking.
• Reduced inference latency by ~30% by converting PyTorch models to ONNX, enabling real-time processing at 20–25 FPS on edge-ready deployment setups.
• Developed weighted-averaging logic and advanced augmentation strategies to preserve high detection reliability in low-light, high-speed, and partially obstructed scenarios.
• Published the research work at Springer; publication link: View publication.