Bassem Ben Hamed

Give analysts an assistant able to answer complex, multi-document questions over a business corpus (annual reports, issue prospectuses, sector research notes, credit agreements, BCT/CMF regulatory texts). A naive RAG falls short: the questions require multi-hop reasoning (linking an issuer, its subsidiaries, its collateral and the applicable regulation).

The LLM is exposed through an OpenAI-compatible API, allowing a switch between cloud and on-premise deployment depending on confidentiality constraints.

• Graph knowledge base (Neo4j) — LLM-based extraction of entities and relations (issuers, subsidiaries, instruments, collateral, executives, regulatory texts) into a knowledge graph queried in Cypher; every node stays linked to its source chunks for citation.
• Hybrid retrieval (GraphRAG) — Neo4j native vector index, full-text search (BM25) and graph traversal.
• Agentic layer (LangGraph) — router agent performing query decomposition and rewriting, then selecting the retrieval strategy (vector, LLM-generated Cypher, or hybrid).
• Advanced RAG techniques — cross-encoder reranking, self-assessed corrective loop (Corrective RAG) and groundedness check before returning an answer.
• Guardrails — answers always sourced (document, page, graph node), explicit refusal when corpus coverage is insufficient.
• Evaluation — RAGAS metrics (faithfulness, answer relevancy, context precision) against the baseline vector RAG, on a question set validated by the analysts.

A real-time view of operational activity (travel files, parcels, Loaded/Unloaded statuses), fed by extraction from SQL Server.

Fully containerised on Kubernetes. KPIs are recomputed cumulatively after each micro-batch, keeping the stream and the dashboard consistent.

• Apache Kafka (KRaft) — ingestion of the vectorys.dossier and vectorys.dossier_sub streams.
• Spark Structured Streaming — 10 s micro-batches: parsing, deduplication, enrichment, aggregation.
• Apache Iceberg on MinIO (S3) — transactional storage through a REST catalogue.
• Trino — SQL query engine; Streamlit — KPI dashboard auto-refreshed every 10 s.
• Apache Airflow (KubernetesExecutor) — controlled start/stop of the producer and the Spark job, health sensors (Kafka, MinIO, Trino), Iceberg maintenance (compaction, snapshot expiry), data quality checks, automated replay and alerting.

Manual interpretation of coronary angiograms is time-consuming and subject to inter-operator variability. An end-to-end AI pipeline, from video to generated medical report, as a clinical decision support tool.

A working end-to-end proof of concept validated with the department's cardiologists; the work was published in international conferences and a journal, with award-winning posters at medical congresses.

• AngioGet — coronary artery segmentation with U-Net.
• AngioKey — extraction of the most informative keyframes from the video sequence.
• AngioID — anatomical classification of the coronary tree (CNN, InceptionV3 fine-tuning, dual-stream).
• AngioView — recognition of acquisition angles and mapping to standard clinical views.
• AngioQ — stenosis detection and grading (YOLOv8, benchmarked against YOLOv1 and Vision Transformer).
• AngioIQR — generation of a structured medical report with Transformers (GPT, BERT), then quality assessment of the generated reports.
• DataCath — a new dataset annotated by the department's cardiologists through a dedicated web annotation platform, released for research.

A shoplifting-detection product was crashing after roughly two hours with a GPU out of memory error, and its code — organised as blocks of scripts — was hard to maintain and deploy.

Outcome: no more memory-saturation interruptions, continuous inference service, and a maintainable code base ready to scale to multiple cameras.

• Memory diagnosis — inference profiling and leak hunting on the Python side (tracemalloc) and the GPU side: TensorFlow allocations were not released between inference cycles.
• GPU allocation optimisation — explicit periodic release (tf.keras.backend.clear_session(), graph reset), per-process memory growth limits, and per-GPU workload sizing based on co-resident models.
• Refactoring into an API — a FastAPI/Uvicorn service with one endpoint per pipeline stage (preprocessing, object detection, action detection), separating model lifecycle from application logic.
• Operational resilience — process supervision with supervisord, with automatic restart as a last resort.
• Knowledge transfer — code review with the client team, documented diagnostic and operating procedures.

Demonstrate industrialised ML model deployment relying exclusively on open source tooling, from raw data to a served model.

A complete demonstrator handed over to the client as an open source reference architecture.

• Development and data — Jupyter; data and artefact versioning with DVC, pipeline sharing through DagsHub.
• Experiment tracking — MLflow (tracking, model version comparison, model registry).
• Quality and testing — Deepchecks (drift, integrity, performance) wired into the pipeline.
• CI/CD — Jenkins: on every change, run tests, rebuild images and redeploy.
• Serving — FastAPI REST prediction API and a Streamlit demo application, both containerised (Docker/DockerHub).
• Deployment — Kubernetes: scaling, rolling updates, high availability.
• Monitoring — Grafana (infrastructure, alerting) and Arize (prediction drift, incoming data quality).

Coaching of the internal data science team on customer intelligence, from business framing through to models.

Outcome: a team autonomous across the full approach, segments used for campaign targeting, and LTV scores feeding the prioritisation of retention actions.

• Segmentation and profiling — behavioural features built with the RFM approach (Recency, Frequency, Monetary) from consumption and top-up data, then K-Prototypes clustering handling numerical and categorical variables jointly (plan type, technology, geography); cluster count selection (elbow, silhouette) and translation of each segment into an actionable business profile.
• Cross / up / down-selling — per-offer propensity scoring models using RFM segments and subscription history.
• LTV — lifetime value modelling to prioritise retention actions and arbitrate acquisition costs.
• Coaching — joint working sessions on notebooks, code and methodology reviews.