Adnane El Mansouri

Designed and deployed a multi-agent conversational AI system to assist client advisors in navigating POS systems and retrieving client data. In parallel, built advanced forecasting pipelines for luxury products across multiple time horizons and granularities to optimize inventory and strategic planning.

- Architected a multi-agent orchestration system using LangGraph, with specialized agents for POS navigation, client data retrieval, analytics and product lookup coordinated via a supervisor agent for dynamic task routing.

- Integrated external data sources and action delegation through MCP (Model Context Protocol) enabling real-time use and structured access of the CRM, inventory, and POS systems.

- Built a RAG pipeline with LangChain combining vector search and knowledge graphs over POS manuals, client data, and product catalogs for grounded, context-aware responses.

- Implemented agentic tool-use patterns allowing agents to autonomously call forecasting models, query BigQuery, and trigger actions based on client segmentation outputs.

- Prompt Engineering of GPT and Gemini as backbone LLMs, and guardrails for reliable advisor-interactions.

- Built demand forecasting models using classical methods (ARIMA, SARIMA), deep learning (LSTM, Temporal Fusion Transformer), and Prophet — with hyperparameter tuning via Optuna and AutoML with PyCaret.

- Enabled multi-horizon, multi-granularity forecasting across product categories and regions, feeding outputs into agent-based decision workflows for supply chain and advisor teams.

- ML pipeline deployment with CI/CD for continuous model retrain and deployment on Cloud Run

Tech Stack: Python, LangGraph, LangChain, MCP Servers, RAG, GPT, Gemini, Multi-Agent Orchestration, ARIMA, SARIMA, LSTM, Prophet, Temporal Fusion Transformer, BigQuery ML, Vertex AI, PyCaret, Optuna, GCP (Cloud Run, Dataflow, BigQuery, Cloud Composer), CI/CD.

POC of an advanced medical recommendation engine, aimed at reducing the time of medical interactions. This system uses NLP and LLM techniques to build KPIs that represent the patient's detailed health profile through various data sources (Medical reports, analysis reports, survey responses, etc.) and on the other hand, issue recommendations on actions to take based on these KPIs such as medical products or changes in lifestyle.

Achievements:

- Implementation of large-scale language models (multimodal LLM) like GPT, Llama, Mistral for tasks like information extraction, medical data analysis (text & image), and recommendation generation.

- Implementation, Prompt Engineering, RAG of LLMs via Langchain.

- Configuration and management of data environments on AWS Cloud, notably AWS S3 and AWS Redshift.

- Deployment of LLMs use cases as end points via AWS SageMaker.

- Setting up data ingestion and processing pipelines on Databricks via Spark.

- Orchestration of different prediction and reporting pipelines on Databricks in batch.

- Maintenance of CI/CD processes to ensure continuous updating and deployment of models and applications.

Tech Skills & Stack: Python, NLP, LLM, GenAI, Langchain, QLora, Prompt Engineering, RAG, transformers, Databricks, Chainlit , MLOps, AWS Cloud, AWS SageMaker, AWS CloudFormation, Amazon CloudWatch, AWS S3, AWS Redshift, Spark, CI/CD.

My intervention focused on setting up a sales forecasting pipeline for dairy products (Demand Forecasting). The project aimed to integrate reporting for supply chain managers from different Time Series modelings to optimize inventory & price management and anticipate market trends, thus dynamically adapting to the fluctuating needs of customers.

Achievements:

- Design and implementation of a machine learning sales forecasting pipeline using advanced Time Series Forecasting techniques over different time horizons and granularities. (ARIMA, Prophet, Boosting models)

- Incorporating deep learning models like transformers architectures for probabilistic forecasting.

- Interpretability of the forecasts with business and different stakeholders.

- Deployment and management of machine learning pipelines on Databricks. (MLOps)

- Setting up data ingestion and processing pipelines via Spark on Databricks and making them available on Snowflake.

- Management of continuous integration and delivery (CI/CD) with Gitlab and Azure DevOps.

Tech Skills & Stack: Time Series Forecasting, Machine Learning, Python, SQL, Azure Cloud, Azure DevOps, Azure Data Lake, Databricks, Spark, Snowflake, SQL, Gitlab, CI/CD.