AI Engineer &Actuarial Data Scientist

Forex markets require analysis of multiple data sources including market prices, macroeconomic indicators, and sentiment data. Traditional single-model approaches fail to capture complex interdependencies.

Built a multi-agent AI system using LangGraph where specialized agents analyze different data modalities and collaborate through a reasoning framework to generate trading signals.

LangGraph multi-agent architecture with specialized agents for market data, macroeconomic analysis, sentiment processing, and signal generation. Uses RAG for context-aware decision making.

Successfully integrated 3 data modalities, achieved explainable trading signals with SHAP-based feature importance, and demonstrated improved signal quality over baseline models.

Asset valuation requires extracting data from diverse documents (PDFs, Excel) and applying complex financial models. Manual processing is error-prone and time-consuming.

Developed an end-to-end AI platform combining document AI for data extraction, LLM-based understanding, and automated financial modeling for DVF valuation, insurance value, and replacement cost estimation.

Document processing pipeline with OCR and LLM extraction, financial modeling engine, and web interface for automated report generation.

Automated 80% of manual valuation work, reduced processing time by 70%, and improved accuracy through standardized extraction and modeling.

Traditional actuarial models (GLM) provide point estimates without uncertainty quantification. Deep learning models lack interpretability and calibration for insurance applications.

Research platform combining classical actuarial methods (GLM) with deep learning (CANN) and uncertainty quantification (NGBoost) for pricing, reserving, and fraud detection.

Modular architecture supporting multiple modeling approaches with conformal prediction for uncertainty bounds and SHAP for explainability.

Demonstrated improved calibration over traditional models, provided uncertainty estimates for risk management, and achieved comparable accuracy with better interpretability.

Life insurance pricing and reserving require accurate mortality forecasting. Traditional static models fail to capture mortality trends and improvements over time.

Implemented stochastic mortality models (StMoMo) using Human Mortality Database data to forecast mortality rates and value annuity products.

Statistical modeling framework with multiple mortality models, demographic data integration, and annuity valuation engine.

Successfully modeled mortality improvements across multiple populations, generated 10-year mortality forecasts, and produced annuity valuations with confidence intervals.

Property-Casualty reinsurers price catastrophe risk using vendor cat models recalibrated on 3-5 year cycles, treated as black boxes, and increasingly disconnected from the pace of climate change. Actuaries spend enormous manual effort translating model output into underwriting decisions, Solvency II reports, and treaty pricing memos.

Production-grade platform that continuously ingests physical climate data alongside exposure, claims, and financial-market data to produce a dynamically updating, climate-adjusted view of catastrophe risk. Uses an ensemble of specialized ML models (Gradient Boosting, Temporal Fusion Transformer, Graph Neural Network, Vision Transformer) fused into a portfolio loss distribution engine, wrapped in a multi-agent LLM system for autonomous report generation.

Azure-native end-to-end platform with Azure Data Factory for ingestion, Azure Data Lake Storage (Bronze/Silver/Gold), Azure Databricks for feature engineering, Azure ML for model training/registry, Azure OpenAI for multi-agent LLM orchestration with RAG over treaty wordings and Solvency II regulation, AKS for FastAPI serving, and React/Next.js + Power BI frontend.

Designed a full MLOps/LLMOps production system with CI/CD, drift detection, automatic retraining, and explainability layer (SHAP/LIME/Counterfactuals). The multi-agent system autonomously drafts treaty pricing memos, answers regulatory questions, runs natural-language "what-if" stress tests, and produces board-ready reports with full citation trail.