Service 02 · AI & Machine Learning
Models are the easy part.
The hard part is the data underneath, the systems either side, and the question of whether anyone can tell that the answer was right. Noble builds the whole of that, not the demonstration.
A surface under continuous computation — the service’s visual, running live where the device can afford it.
A model reads what the asker may read
A retrieval system inherits every permission mistake in the content it is allowed to read. So the boundary sits before the ranking, not after it: a highly relevant document outside this asker’s permission is not retrieved — which is exactly where systems that rank first and ask later go wrong.
What this service is
Applied engineering on top of models Noble does not own. The models come from the market and change every few months; what does not change is the work of getting an organisation’s own data into a state where a model can be trusted against it, and of putting the result somewhere a person can act on.
That is why this service sits beside Data and Software Engineering rather than above them. An assistant with nothing reliable to read is a demonstration; the same assistant over a governed data layer is a system.
The service, by capability family
Eight families, grouped by the shape of the problem rather than by whichever technique is current. Each is named where Noble can build it, integrate it with the systems around it, and keep it running afterwards.
AI advisory and architecture
- Use-case discovery, and saying which ones are not worth building
- AI architecture against the systems already in place
- Readiness assessment — data, access and operating capacity
- Model and platform selection, revisited as the market moves
- Adoption planning, because an unused system has no value
Generative AI and LLM applications
- Generative AI applications built on large language models
- Retrieval-augmented generation over an organisation’s own content
- Enterprise assistants and copilots inside existing applications
- Knowledge systems built on governed, permissioned sources
- Prompt, context and retrieval design treated as engineering
AI agents and intelligent automation
- Agents that call real systems, with bounded permissions
- Tool use and orchestration across applications and APIs
- Agentic workflows with a defined stopping condition
- Process automation where the rules are stable
- Human review where the cost of being wrong is high
Machine learning and predictive intelligence
- Forecasting against operational history
- Anomaly and exception detection
- Classification and scoring models
- Recommendation and ranking
- Decision intelligence — scenarios compared under real constraints, for a person to choose between
Computer vision and document intelligence
- Image classification and object detection
- Visual inspection against defined criteria
- OCR and extraction from scanned and photographed documents
- Document understanding — structure, tables and relationships
- Multimodal processing where text and image arrive together
Conversational, voice and multilingual AI
- Conversational interfaces over real systems, not scripted trees
- Voice interfaces, transcription and spoken interaction
- Arabic and multilingual handling as a first concern, not a translation layer
- Natural language processing — classification, summarisation and search
- Enterprise search across content people are permitted to see
AI engineering and MLOps
- Data pipelines and feature preparation
- MLOps — model serving, versioning and rollback
- Evaluation sets built before deployment, not after
- Monitoring for drift and silent degradation
- Cost and latency treated as design constraints
- Integration into the systems people already use, the ERP included
Responsible AI, security and governance
- Access control over what a model may read
- Personal data handling and residency decisions
- Audit trails over prompts, retrievals and actions
- Evaluation and human oversight proportionate to the risk
- Model lifecycle governance — what changed, when, and who approved it
- Documented limits, and a defined path when the model is wrong
The service, in depth
Seven specialist pages beneath the families above. Each describes a system rather than a capability — what it reads, what it may do, how it is evaluated, and who holds the boundary.
- Generative AI & LLM ApplicationsAssistants, copilots and knowledge systems on large language models — grounded in an organisation’s own permissioned content, and evaluated before anyone is shown a demonstration.
- Agents & Intelligent AutomationAgents that call real systems on bounded permissions, with a defined stopping condition and a person in the loop wherever being wrong is expensive — and process automation where the rules are stable enough not to need one.
- Machine Learning & Predictive IntelligenceForecasting, anomaly detection, classification, recommendation and decision intelligence on an organisation’s operational history — evaluated on held-back data before deployment, and advisory until the record says otherwise.
- Computer Vision & Document IntelligenceImages, scans, photographs and camera streams turned into structured records with a confidence attached — inspection against defined criteria, extraction from Arabic and English documents, and a person for whatever the model is unsure of.
- Conversational, Voice & MultilingualConversational interfaces over real systems, voice where the hands are busy, and Arabic handled as a first concern — dialect, code-switching and right-to-left composition — rather than as a translation layer over an English design.
- AI for Enterprise SystemsHow a model reaches the ERP, the CRM, the service desk and the document store without becoming a dependency on their internals — released interfaces, a bounded identity, governed data and a record of every read and write.
- Engineering, MLOps & GovernanceThe pipelines, serving, versioning, evaluation and monitoring that keep a model honest after it ships — and the governance that can say what changed, when, who approved it, and what the model is not allowed to do.
From a question to something in production
- 01
Frame
What decision changes if this works, and what would count as being wrong. A project that cannot answer the second question is not ready for the first.
- 02
Data
Find it, join it, and establish what it can and cannot support. Most of the effort on most projects is here.
- 03
Build
The smallest system that answers the question, against the evaluation set rather than against a demonstration.
- 04
Evaluate
Measured on held-back data, including the cases it is expected to fail on. The result is a decision about whether to deploy at all.
- 05
Operate
Serving, monitoring, and re-evaluation as the data underneath moves. A model is not finished when it ships.
What it depends on
This service is unusually dependent on the others. These are the ones an AI project reaches into before it reaches a model.
- Data & AnalyticsNothing here works on data that disagrees with itself. This is the dependency, not one of several.
- Software & Web EngineeringA model reaches a person through an application. Somebody has to build and maintain that application.
- SAPOn an enterprise estate the operational record usually lives in SAP, which is where prediction and document work find something real to run on.
- CloudWhere the workload runs, what it costs, and which region the data is allowed to be in.
- SecurityRetrieval is an access-control problem before it is a ranking one, and the permissions it enforces have to come from the same identity system as everything else.
Where to start
With a decision somebody makes repeatedly, and a record of how it has been made before. That is enough to establish whether there is a project here at all, and it is a cheaper question to answer than most organisations expect.