Jean-Michel CARUGE

• Built gradient boosted trees classifiers as web application firewalls to protect Akamai's customers from malicious web traffic. Achieved false positive rate below 0.5%. Collaborated with engineering team to (1) deploy models in production and (2) track KPIs in real time. Model stack is n-gram extraction + vector embedding + principal component analysis + GBT classifier.
• Built a ML based, spark streaming application to detect distributed denial of service (DDoS) cyber-attacks. It sends detection signals to the Security Operations Control Center, which mitigates the DDoS attack. Detected 99.5 % of SIN flood and 85 % of DNS flood. Spark MLLib implementation (PySpark/SQL) with Azure Event Hubs streaming and Azure Storage.
• Pioneered web session tracking (instead of signature tracking) to identify malicious agentic web traffic. Word2Vec embedding with large context window uncovers correlation between sequences of web page visits. Detected google-shopping, chatgpt-user, comet browser with more than 93.5 % accuracy.
• Crafted prompt engineering with LLMs in MCP framework to discover/fixed vulnerabilities in cyber-defenses (red teaming).
• Reduced third party cloud spend by $65,000 per month by (1) migrating daily ML jobs on-premises and (2) balancing small size ML models with KPIs constraints.
• Mentored junior data scientists on advanced statistical techniques and best practices.

• Trained NLP, NLU and Deep Learning models for Amazon Alexa. Added/Fixed more than 130 features for Alexa, FireTV and Echo Show working in a CI/CD DevOps environment.
• More than 20 on-calls. Provided technical expertise during production-related severity incidents: root cause analysis, quick model fix, and deployment.
• Proficiency with exact matched and pattern matched rules, finite state transducers, NLP, NLU, conditional random fields and automatic speech recognition models.
• Mentored junior data scientists, fostering skills in advanced analytics and model development.
• Automated repetitive tasks using scripting languages such as Python, Shell Scripting, saving time during the analytical process significantly.

• Built real-time trading algorithms for sovereign bond traders. Between 5-7% positive impact on P&L, which is significant since top traders can have multi-million-dollar daily profit.
• Migrated legacy ExcelVBA trading algorithms onto modern C# trading platforms.

• Use random forest and Xgboost ML models to optimize database marketing. Measured a 20 % improvement in customer acquisition, click rate and customer conversion rate.

Managed a cross-functional team to develop quantum dots LEDs for displays.

Managed budget of $500,000 and 6 full time employees.

Planned and scheduled deadlines for 2 projects per month.

• Modeled financial market risks and calibrated stochastic models for real-world and market-consistent scenarios. Used SDEs, correlation matrix to correlate stochastic shocks, and Monte Carlo to project cashflows for value at risk, conditional tale expectation projections.
• Provide technical expertise for sales teams during the acquisition of multi-million-dollar contracts.