saber amri

• AI Document Intelligence PlatformAt GAN Assurances, I designed and deployed a modular multi-agent system (9 agents) that automates document analysis across claims and back-office workflows, with the goal of speeding up claim processing and reducing manual workload for claims handlers. • Architecture: microservices orchestrated by LangGraph. Each agent (image quality, preprocessing, orientation, layout analysis, OCR, structuring, NER, decision) is an independent FastAPI microservice, deployed and scaled on its own. Orchestration is handled by LangGraph rather than a hand-written sequence of calls: every node of the state graph calls one microservice over HTTP and writes its result into a typed shared state. A central router first classifies each incoming item: documents go to the extraction graph (to be stored, or analyzed and matched against the claim's missing documents); questions go either to a direct claims-database lookup or to the RAG engine. Conditional edges make every routing decision explicit and visible in the graph: rejection of unusable images, empty layouts, low-confidence OCR, agent failures. Extraction (agents 1 to 7) and decision-making (RAG, LLM, Next Best Action) are decoupled, so each side can evolve and scale independently. • Loop engineering: recover instead of spinning. Every loop in the pipeline is bounded and observable: 1- Resilience loops: per-node retry policies with exponential backoff for network errors and transient agent failures; only persistent failures reach a terminal failed state, along with the faulty step. 2- Schema-validation loops: LLM outputs are parsed into Pydantic schemas (structured outputs); an invalid or incomplete output triggers a bounded regeneration. 3- Feedback loops: downstream agents flag weak upstream results and trigger a deeper pass. Low OCR confidence sends the document back to preprocessing with stronger enhancement before a new OCR pass; an incomplete or unsupported answer triggers retrieval widening, targeted retrieval of missing points, then escalation to a stronger model. 4- Human-in-the-loop: graph state is checkpointed after each node, so an interrupted run resumes where it stopped, and a document can be paused for validation by a claims handler, then resumed with the corrected data. • RAG engine for Next Best Action. The engine (GPT-4, Mistral, Claude) recommends the next best action for each claim, with cited sources. Retrieval combines semantic chunking with adaptive overlap, metadata filtering, hybrid dense + BM25 search and cross-encoder reranking. A deterministic dispatcher decides whether the top-1 passage is enough (sequential reading for single-value questions) or whether the top-K is needed (batch reading for lists and comparisons). Answers carry completeness and confidence flags that drive the widening and escalation loops. Business rules (priority, SLA, required documents) stay in a deterministic NBA engine, and the LLM explains and documents the recommendation. • Prompt engineering as an engineering discipline. Prompts are versioned, outputs are typed (Pydantic), negative constraints remove hedging and enforce an exact refusal when information is missing, and sources are mandatory. A CI pipeline compares prompt variants on evaluation sets (RAGAS metrics: faithfulness, answer relevancy, context recall and precision) with a development / test split, and blocks any deployment that regresses. • Workload management. Incoming documents go through a queue (Celery, with Redis and RabbitMQ): retries, a dead-letter queue for repeated failures, and workers scaled dynamically on queue depth. Each worker runs one LangGraph graph per document, so document batches absorb peak volumes without degrading response times. • Industrialization and observability. GitLab CI/CD covers automated tests, Docker builds and registry pushes. The FastAPI services and AI models are containerized and deployed on Kubernetes (Deployments, Services, HPA autoscaling). MLflow tracks experiments and model versions in production. Monitoring covers metrics, logs and traces (Prometheus, Grafana, Loki, OpenTelemetry) with alerts on latency, error rates and document rejection rates, plus LLM-level tracing (prompts, latency, cost, scores) to detect quality drift.

Significant experience of 6 months with sas at the Parnasse client. data extraction for a recovery and marketing need

Implementation of a questionnaire (audio interviews) as part of a program called Datadriven intended for employers to improve the communication strategy, measure the quality of employee engagement and develop areas of recommendations, Coding of questionnaire, statistical analysis of data (Unsupervised classification of employees and Recommendation of communication strategies for each class of personae), transcription of audio data into texts (Speech to Text) and Analysis of textual data with NLP.

Convert scripts from R to Python. Loading and transforming data with Spark and Apache Drill.

Assistance with the migration of the Caffe2 neural network framework to be trained and deployed from AWS managed environments via SageMaker, Athena and Amazon S3: Detection of anomalies on metro tracks thanks to the analysis of the video stream from a placed camera on a motor, launch commands via the SageMaker APIs for the subscription of a machine mounted on the fly that retrieves the data in S3 and performs the training: SageMaker has since made it possible to monitor the training activities on the machine allocated to the calculation. FEATURE ENGINEERING: Data processing and cleaning, Detection of outliers (extreme value) by eliminating them from data, Implementation of machine learning algorithms to predict the failure, Define indicators and implement failure management tools, Implementation e models of time series (GARCH, ARMA, IGARCH ..) to predict measurement indicators.

Customer knowledge: Classification of customers through RFM Scores, Recommendation engine: Calculation of similarities between customers and products, Association rules and collaborative filtering

Our solutions incorporate face detection technology and gesture recognition to allow brands to better interact with consumers and to enable consumers to live a more enriched user experience. The audience analysis of billboards and collaborative intelligence are the heart of our business. Project: Learn billboards to make the decision. Collecting traffic and audience data by manipulating data stored under MongoDB. Improved Dashboard with Qlikview and ClicData tools, featchering and Audience Analysis, auction prediction (Advertising Price) using Machine Learning and Deep Learning models, Microsoft Azure Test and building time series models capable of present the predictions (The goal is to send alerts to customers in case there is a high traffic or audience). Test models on store data of gas stations: The goal is to predict the audience.

Creativ-Ceutical provides strategic pricing, market access and health economics consulting in life sciences. Project: Statistical study of data from the Program of Medicalization of Information Systems PMSI, Identifications of hospitalizations related to breast cancer and comparison of its management between private and public institutions (SAS software).

Analyst Dataminer: Data reporting with plsql (loader, shell and batch techniques to automate data loading) universe design and BO reporting development. Building profitability prediction models for bank customers: machine learning models such as Random forst, Bagging, Boosting, SVM, Neurol network...