Global Summit on Data Science and Cloud Computing

Attendees, exhibitors & solutions

Who this event is for

Industry trends

Key trends shaping this market and its wider industry.

Hotel tiers

A range of accommodation options near the venue and across key areas.

Local transport

Get around the city with these convenient transport options.

Frequently asked

Glossary

1

MLOps

The operating discipline that keeps machine-learning models running in production: versioned training data, reproducible experiments, automated deployment, drift monitoring and rollback. At a two-day data science conference the MLOps track is where platform engineers compare orchestration tools and ask how small teams keep models auditable without a dedicated platform squad.

2

Feature store

A shared layer that computes and serves model inputs consistently for both training and live inference, avoiding train/serve skew. Sessions on feature stores usually contrast a centralised platform with domain-owned features, a trade-off that matters for banks and telcos with many separate data teams.

3

Retrieval-augmented generation (RAG)

An architecture where a language model answers using documents retrieved at query time rather than relying only on its training weights. RAG is the dominant pattern for enterprise assistants because it lets teams update knowledge without retraining, but answer quality depends on chunking strategy, embedding model choice and retrieval evaluation.

4

Vector database

A store optimised for similarity search over embeddings, used to power semantic search, recommendation and RAG retrieval. Conference discussion typically covers hybrid search (dense plus keyword), index rebuild cost, and whether a general-purpose database with vector extensions is good enough for the team's scale.

5

Fine-tuning versus prompting

Two ways to adapt a foundation model: instruction or prompt engineering with a frozen model, or weight updates on domain data. Fine-tuning buys behaviour and tone control but adds hosting, evaluation and versioning burden, so the practical question at this conference is which tasks justify the extra machinery.

6

Data lakehouse

An architecture that combines cheap object storage with table formats providing ACID transactions and schema enforcement, so BI and ML read the same governed tables. Adoption talk centres on open table formats, catalog interoperability and migration cost from a legacy warehouse.

7

Data mesh

An operating model that treats data as a product owned by domain teams, with a central platform providing self-service infrastructure and federated governance. It is an organisational answer rather than a tool, so conference discussions tend to focus on ownership contracts, discovery catalogs and how to keep quality measurable.

8

Model governance and the EU AI Act

The policies, documentation and controls that evidence how a model was built, validated and monitored, mapped to risk tiers under the EU AI Act. Teams working in Europe increasingly need model cards, dataset lineage and human-oversight records that satisfy both internal audit and regulators.

9

Synthetic data

Artificially generated records that preserve statistical properties of real data without exposing personal information. It is used to unblock rare-event modelling, testing pipelines and privacy-sensitive sharing, but validity checks against the real distribution remain essential before production use.

10

Federated learning

Training a shared model across separate data holders who keep their raw records local, exchanging model updates instead. Relevant where data cannot leave a hospital, bank or device, and the recurring engineering questions are communication cost, straggler handling and privacy guarantees.

11

AutoML

Automated search over preprocessing, model families and hyperparameters to produce a baseline faster than manual iteration. In practice teams use it for breadth and benchmarking, then hand-tune the few pipelines that reach production, which is how it is usually framed in applied sessions.

12

Model drift

Degradation in model performance when live data diverges from training data, detected through distribution shift metrics and delayed outcome labels. Monitoring design, alert thresholds and retraining triggers are standard MLOps content because most production failures appear here first.

13

Evaluation harness

A repeatable test suite for model quality, covering task accuracy, safety behaviour, latency and cost per request. For LLM applications the harness usually blends automated scoring with curated human review sets, and it is the artefact that makes model changes releasable rather than improvised.

14

Responsible AI

The practice of documenting and mitigating bias, privacy and misuse risks across the model lifecycle. At a European conference it usually appears bundled with governance obligations: fairness testing, impact assessments and clear escalation paths when a model behaves outside its declared scope.

15

Data quality contract

An explicit, testable agreement between a data producer and its consumers covering schema, freshness, completeness and null thresholds. Contracts turn data incidents into detectable failures rather than silent model degradation, which is why they recur in lakehouse and data mesh sessions.