Designing AI agents for regulated environments

Joining BCG X to design and ship production AI and agentic systems for enterprise clients across regulated industries.

• Design and deliver production agentic AI systems for enterprise clients, from prototype to governed deployment.
• Partner with cross-functional teams to embed LLM copilots into regulated business workflows with evaluation and guardrails.
• Scale applied GenAI solutions with a focus on reliability, compliance, and measurable business impact.

Leading architectural initiatives for next-generation AI integration protocols and multi-agent coordination patterns within enterprise banking infrastructure.

• Architecting Model Context Protocol (MCP) design patterns enabling seamless AI agent orchestration across distributed banking systems.
• Driving proof-of-concept implementations for the Agent-to-Agent (A2A) protocol, establishing communication standards for autonomous service coordination.
• Delivering technical blueprints for scalable AI infrastructure supporting compliance-first financial services environments.

Architected a production text-to-SQL assistant empowering 40+ financial analysts to interrogate complex datasets without SQL expertise.

• Architected a Text-to-SQL system enabling 40+ financial analysts to query complex financial databases in natural language, removing the IT-team bottleneck.
• Semantic parsing + query optimization hit 85.6% accuracy and cut average response time from minutes to under 20 seconds.
• Built an evaluation framework on accuracy, latency, and user satisfaction; decision-making cycle reduced by 27% thanks to faster insights.
• Delivered production-ready AWS pipelines with scalability and financial-services compliance baked in.

Brought LLM reasoning to vehicle telemetry, diagnosing anomalies for connected cars within the R&D division.

• Designed an LLM-powered diagnostic system analyzing connected-vehicle sensor streams to detect and explain anomalies in natural language.
• Shipped a knowledge-retrieval pipeline linking anomalies to troubleshooting docs and repair procedures, with a root-cause layer translating sensor data into actionable repair insights.
• Achieved 72.8% anomaly detection accuracy, reduced diagnostic time by 24.5%, cut misdiagnosis by 14.2%, an estimated €1.1M/year saved in warranty claims.

Partnered with the TER division to curb multimillion-euro ticket fraud through predictive controls.

• Inside TER's data squad, brought ticketing fraud rate from 7.2% → 5.9% through hypothesis testing and optimized controls.
• Developed predictive models (RF, XGBoost, LSTM) with exogenous variables, achieving AUC-ROC > 0.87.
• Regression models forecast fraud volumes at 14.7% RMSE, enabling smarter inspector resource allocation.

Advanced operational safety initiatives for France's gas transport network within a 6-person innovation program.

• Gas Propagation Tool: built predictive time-series models for pipeline gas dynamics, improving incident anticipation across France's gas transport network.
• LLM Chatbot: led development of a Confluence-integrated assistant, reducing support response times by 38.5%.
• Introduced TDD and continuous retraining discipline across the team, materially improving maintainability.

Scaled AI products for thousands of legal professionals within a dedicated AI Operations squad.

• Stood up a full MLOps platform: deployment cycles down 41%, release reliability materially improved.
• Refactored Python codebases to cut tech debt and built CI/CD pipelines (GitLab, ZenML, MLflow, DVC) for automated ML workflows.
• Managed AWS infrastructure with Terraform (incl. SageMaker) for scalable deployment of ML products used by thousands of legal professionals.

Strengthened fraud prevention and AML/CTF compliance inside the Risk & Compliance department.

• Implemented an AML/CTF fraud detection system, improving compliance posture and reducing false negatives.
• Built PySpark/Hadoop pipelines lifting data quality across millions of transactions; automated workflows cut manual workload by 24%.
• Leveraged Neo4j (CypherQL) to detect fraud networks and suspicious relational patterns.

Automated document analysis for innovation labs by combining OCR and NLP workflows.

• Built ML/DL models for OCR of printed and handwritten text, cutting manual processing time by 34%.
• Hyperparameter tuning lifted prediction accuracy by +11.8% F1-score; ran text-similarity analysis for document classification and deduplication.

Supported modernization of citizen-facing digital services with NLP automation.

• Built NLP models for text classification, topic modeling, and chatbots supporting modernization of citizen-facing digital services.
• Automated information-extraction pipelines and deployed Flask web apps + scraping pipelines in production.