LLM
Large language models
Applications built on language models where the value is in what surrounds the model: what it is allowed to read, how that context is retrieved, how the output is constrained, and what it is permitted to do with the result.
Engineered on request
- Enterprise LLM applications
- Retrieval-augmented systems
- Structured information extraction
- Classification and routing
- Summarisation over operational records
- Structured outputs against a schema
- Tool and system interaction
- Grounded answers with sources
Works with: text & documents · business data
NLP
Natural language processing
Not everything needs a language model. A great deal of business text is better handled by classification, extraction and retrieval that are cheaper, faster and more predictable — and can be evaluated.
Engineered on request
- Text classification
- Entity and field extraction
- Document analysis
- Semantic and keyword retrieval
- Intent understanding
- Text-derived business signals
Works with: text & documents
CV
Computer vision
Vision work is only useful when its output enters a workflow. The engineering is in the pipeline around the model: capture, pre-processing, thresholds, review of uncertain cases, and what happens to the result.
Scoped per engagement
- Image analysis and classification
- Video analysis
- Object detection and recognition
- Visual inspection
- Visual tracking
- Document and image understanding
- Vision-driven business workflows
Works with: images · video
Speech
Speech & audio
Speech recognition and audio processing engineered into workflows where typing is impractical — a warehouse, a vehicle, a production floor — and where the transcript then has to reach a system that can use it.
Scoped per engagement
- Speech recognition and transcription
- Spoken-language interfaces
- Audio processing
- Voice-enabled workflows
- Hands-free field operation
Works with: speech & audio · text & documents
Multimodal
Multimodal systems
The point at which these stop being separate technologies. A delivery exception may involve an ERP record, a photograph, a driver’s spoken note, a GPS position and a signed document — and the useful system is the one that considers them together.
Scoped per engagement
- Combining records, documents, images and audio in one workflow
- Cross-modal correlation
- Evidence assembled from several sources
- Location context joined to operational data
Works with: text & documents · images · video · speech & audio · business data · location
Interfaces
Intelligence interfaces
The interaction may be conversational. What makes it an intelligence interface is what happens underneath: intent recognised, connected systems queried, sources correlated, findings explained, an action proposed — and, for anything consequential, a person asked first.
Built and running
- Intent and context recognition
- Querying connected systems
- Correlating multiple sources
- Detecting risk, anomaly and opportunity
- Explaining a finding with its evidence
- Proposing actions
- Triggering controlled workflows
- Human approval for consequential actions
Works with: text & documents · business data
Decision support
Decision support
Traditional reporting answers a question that was already formed. Decision support runs continuously against operational data and raises the exceptions — which is a different engineering problem, and a different product.
Built and running
- Continuous evaluation against operational data
- Anomaly and pattern detection
- Threshold and rule evaluation alongside model output
- Prioritised exceptions
- Explanations attached to every finding
Works with: business data · text & documents