Capability 03
AI and data systems
Retrieval, classification, drafting and analysis applied to your own data, with sources, limits and human review designed in.
Problem and opportunity
Useful AI is mostly data work and governance
Model access is commodity. The value comes from clean inputs, a defined task, traceable sources and a review step that a responsible person is willing to sign.
- Knowledge sits in documents nobody can search quickly.
- Generic assistants answer confidently without citing your material.
- There is no agreed boundary for what a model may decide alone.
What Hapis delivers
Scope of delivery
Data preparation
Collecting, cleaning and structuring the documents and records a system will rely on.
Retrieval and search
Question answering grounded in your own material, with the source shown next to the answer.
Extraction and classification
Turning unstructured documents into structured fields your systems can use.
Review and governance
Confidence handling, escalation paths and logs so decisions can be checked.
How the system works
The layers behind the interface
Each layer has a defined responsibility, so the system can be extended without rewriting what already works.
- 01
Source
Defined inputs with owners, refresh rules and access boundaries.
- 02
Index
Chunking, embedding and metadata so retrieval is precise rather than approximate.
- 03
Reason
Task-specific prompts and tools with explicit limits on scope.
- 04
Verify
Citations, checks and human approval before anything acts on the result.
Integration and architecture
Where it connects
- Model access through a gateway so providers can change without a rewrite.
- Vector and relational storage chosen per workload.
- Server-side execution only, with keys held in environment configuration.
- Clear separation between experimental and production behaviour.
Use cases
Typical applications
Internal knowledge assistant
Staff ask questions and receive answers with the underlying document reference.
Document intake
Quotations, specifications or reports parsed into structured records for review.
Analysis support
Drafting and summarising work products that a specialist then verifies.
These describe the shape of the work. Named client references are published only with written approval.
Delivery process
How a project runs
- 01
Discover
Define the task, the data and the acceptable error profile.
- 02
Design
Pipeline, evaluation method and review workflow.
- 03
Build
Iterate against a fixed evaluation set, not one-off demos.
- 04
Integrate
Connect to the interfaces where the work actually happens.
- 05
Improve
Monitor quality and adjust as data and usage change.
Questions
Answered directly
- Does our data train public models?
- We configure systems so client material is used for the task at hand, and we agree data handling in writing before build.
- How do you measure quality?
- With an evaluation set drawn from your real cases, scored before and after changes.
- Can it act automatically?
- Only where the risk is understood. Higher-impact actions keep a human approval gate.
Next