Soufiane Aazizi

– Developed Agentic ICF system that transforms complex Clinical Study Protocol tables (spanning multiple pages, dozens of columns and rows) into concise ICF Summary Tables - reducing generation time from 1-2 days (manual) to under 5 minutes, leveraging Skills Cards architecture with scoped MCP tool restrictions. – Developed Vision Agentic RAG system with DSPy serving a global team of medical writers; improved retrieval hit-rate@6 from 60% to 85% across thousands of indexed documents and images. – Built custom document parser using Docling to extract tables, figures, and images from complex PDFs, RTF, and DOCX into structured Markdown and metadata; indexed into Weaviate (text) and Google Cloud Storage (thousands of images). – Designed comprehensive evaluation framework assessing parsing quality, retriever performance, anti-hallucination robustness, and answer generation effectiveness. – Re-engineered monolithic application with full streaming architecture, reducing long-running task response time from ~2 minutes to ~4 seconds.

– Led a team of 5 Data Scientists; delivered POC in 2 months and production-ready RAG chatbot with full UI in 3 months, parsing thousands of documents (PDF, PPTX, images) for the audit department. – Designed compound AI architecture with DSPy optimizers: query decomposition, chain-of-thought reasoning, and multi-hop document traversal — improving answer accuracy from 70% to 94%. – Implemented Azure Search reranking and enhanced recursive retrieval with DSPy for dynamic keyword generation; achieved ~4-second streaming response time. – Integrated LangFuse for real-time monitoring, performance evaluation, and feedback collection.

– Designed and implemented QA ChatBot leveraging LlamaIndex RAG and LangChain, covering dozens of FAQ topics with sub-50ms retrieval latency, eliminating manual searches for support agents. – Employed auto-retriever composition on Vespa.ai for enhanced passage retrieval; implemented LangFuse monitoring for performance evaluation and continuous FAQ enrichment.

– Developed personalized learning platforms using LLMs (Claude on Bedrock) and hybrid search (Vespa), serving Wiley's global client base of millions of learners. – Designed automated skill-tagging model covering 8,000+ skills achieving 95% precision, with MLOps pipelines (SageMaker, MLFlow, Lambda, Step Functions). – Built hybrid search engine combining ANN and BM25 with spell correction, autocompletion, and multilingual search; deployed via Docker & Kubernetes with 4 APIs. – Created dynamic skills discovery dashboard with semantic search and job-skills alignment (Dash, Plotly).

– Trained and upskilled data science team on Amazon SageMaker and PySpark through hands-on sessions, enabling advanced data analysis and model development capabilities.

– Led KYC remediation using AWS Textract for ID document analysis, MRZ consistency checks via checksum validation, and CRM consistency using text mining. – Built predictive churn risk model using customer banking activity data; optimized with HyperOpt and LightGBM for improved retention strategies. – Advanced email sentiment analysis by developing specialized banking lexicon, adapting VADER method with LLM-derived enrichment (StanfordNLP, NLTK). – Implemented recurrent payments and subscription detection within customer transactions using LLM techniques with PySpark/Hive.

– Developed personalized homepage allocation by transforming unsupervised clustering into a supervised recommendation system using browsing histories. – Created comprehensive cross-sell recommendation system integrating business rules, seasonality, and recurrence; implemented PySpark/HDFS/Hive pipelines.

– Developed and launched Equity Risk Premia (ERP) strategies with full backtesting and production deployment. – Built ERP strategies using ML algorithms (Random Forests, SVM); developed index pricing models (Fear Vol, Value, Quality, multi-factors). – Integrated with trading systems and calculation agents; developed FACTSET API connectors for data mining and ownership analysis.

– Built sales forecasting models for e-learning products; optimized data pipelines on AWS S3/EMR for marketing strategy development.

– Automated data ingestion pipelines and built NLP/text mining prototypes; continuous data recovery via Flume into Hadoop HDFS.

– Research on numerical methods for Backward Stochastic Differential Equations, Lévy processes, and fractional Brownian motion via Malliavin calculus. – Portfolio optimization with stochastic constraints and switching problems.

– Front-office support on Kondor+ and Kondor+ Structured Products for quantitative analysis and pricing.

– OTC derivatives scope including swaps, equity-swaps, CFDs, CDS, swaptions, TRS, and inflation-linked swaps. – Defined pricing processes for hedge and structured fund groups.