Find relevant knowledge
Retrieval-augmented generation, or RAG, brings selected source material into an answer. Where useful, hybrid search combines exact terms—such as product codes—with searches for meaning.
AI / Design, integration, and operation
AI built around how your business works.
SOA designs and implements AI assistants, bounded agents, and applications that use your business information, approved practices, and operational systems. We connect the knowledge each task needs, control what the system can do, and build in ways to review its work.
Client work
An AI platform for recruiting operations.
This ongoing engagement is one example of SOA’s work in AI-enabled business operations. For an external recruiting agency, SOA is building and extending a specialized platform that connects candidate intelligence with discovery, evidence-backed recommendations, outreach, candidate submission, and interview coordination. The client has received a fully functional interface for monitoring AI activity, reviewing its work, and approving candidate outreach, submissions, and interview coordination.
Human-led Platform-led Input Workflow Review & approval
Original workflow
Recruiter-led, using the existing ATS and email.
- Incoming hiring requisitionInput
- HumanSearch and assess candidatesAll candidates sourced by recruiters
- HumanCandidate outreach
- HumanSubmit candidates to client
- HumanSchedule interviews
- HumanFinalize Placement Pending Offer
The requisition is the input. All subsequent steps are recruiter-led, coordinated through the existing applicant tracking system (ATS) and email. All candidates are sourced by recruiters in step 2, Search and assess candidates. The final human step is Finalize Placement Pending Offer.
AI-supported workflow
Platform-led work, with people approving key actions.
- AutomatedDetect incoming requisitionAutomatically
- AutomatedDiscover and assess candidates
- AutomatedRanked recommendationsEvidence, gaps, and uncertainty
- AutomatedCandidate outreachConfirm interest
- AutomatedSubmit candidates to client
- AutomatedCoordinate interviews
- HumanFinalize Placement Pending Offer
AutomatedSupporting capabilityCandidate intelligence
HumanNetwork Provided CandidatesSelective sourcing
HumanHuman review
and approval
and approval
The first six workflow steps and candidate intelligence are platform-led. Candidate intelligence feeds discovery and assessment. Network Provided Candidates is a separate human-led input that feeds directly into Candidate outreach (step 4). Human review and approval apply to outreach and interest confirmation, submission to client, and interview coordination only. The final step, Finalize Placement Pending Offer, is human-led.
The Technology Stack
The platform uses AWS for cloud hosting, Temporal for asynchronous workflows, and Claude Haiku for AI inference. Amazon Nova supports semantic search, while PostgreSQL provides lexical search.
Where we help
Put useful AI into everyday work.
Help staff find an approved answer, extract information from a document, prepare a response, or complete an authorized step in an existing workflow.
SOA assesses opportunities, designs and builds applications, connects systems, evaluates performance, and prepares the people and procedures needed to operate them. The work can begin with a new application, a stalled pilot, or a specific improvement to something already in use.
We build on suitable systems you already have, including document repositories, business applications, identity services, and approval processes. Scope reflects the task, the condition of your information, your budget, and the capacity available to maintain the result.
Business context
Make business knowledge available when the task needs it.
Useful AI needs more than access to a folder. It needs the right information, in a usable form, with a clear basis for deciding which sources apply.
We work with the people responsible for the process to identify authoritative information and capture knowledge that is not yet documented. Inputs can include approved procedures, product guidance, reviewed resolved cases, decision criteria, and feedback accepted by a process owner.
Read documents as part of a workflow
Document intelligence extracts and organizes information from forms, PDFs, and other records, with uncertain fields presented for review.
Check current operational facts
Authorized connections to business systems retrieve live information such as order status, stock availability, or account details.
Apply business rules
Explicit conditions, calculations, and approval limits are enforced through application logic and existing systems. Decisions requiring discretion go to the responsible person.
For each task, the application brings together relevant instructions, supporting information, current records, and permitted tools. This is context engineering: deciding what information the AI receives, when it receives it, and how that information is kept useful.
Source ownership, versions, refresh schedules, deletions, and access rights are part of the design. Corrections and new experience enter a reviewed update process so that changes to guidance and system behaviour can be tested before release.
Agents and workflows
Choose the right amount of automation.
Different tasks need different designs. We choose the simplest design that meets the task and operating requirements.
- Assistant
- Helps someone find information, understand a situation, or prepare work while the person directs the task.
- Predefined workflow
- Follows known steps, such as extracting fields, checking requirements, preparing a draft, and requesting approval.
- Bounded agent
- Selects among permitted tools and next steps when the route varies, within a defined objective and stopping conditions.
These approaches can work together in one application. Connections to business applications use defined tools and limited permissions. Outputs follow agreed formats, required fields and values are validated, and approval points make proposed changes visible before execution.
Access and action limits are enforced by the application and connected systems. Missing information, conflicting instructions, and failed actions have explicit handling paths. The application shows what remains unresolved and routes exceptions to the right person.
Evaluation and traceability
Make the work easy to inspect and correct.
Review should be part of the application people use.
For an answer, show the relevant source passage and document version. For a proposed action, show the record involved, the fields that will change, and any missing information. After execution, distinguish a requested action from one confirmed by the receiving system.
We design records that connect a task to the information retrieved, tool actions, model and configuration versions, and approval history. Users can correct drafts and flag problems; authorized operators can investigate failures and manage changes. Privacy, encryption, access, and retention controls also apply to these records.
Source references and activity records support review. Accuracy and successful execution still need to be tested; these records do not reveal a model’s complete internal reasoning.
Test against the work people need to complete
Evaluation covers representative work and difficult cases: ambiguous requests, missing or outdated information, attempts to obtain unauthorized data, misleading instructions inside documents, and failed or interrupted actions. Acceptance criteria address task completion, answer support, access controls, review effort, response time, and cost.
Keep accepted behaviour under review
After release, monitoring identifies failures and changing performance. Regression evaluation checks whether changes to sources, models, tools, or workflows disrupt previously accepted behaviour. Training, operating procedures, and agreed support responsibilities help the team use and maintain the application.
Model and deployment choices
Choose where the application and models run.
SOA designs around managed model APIs, private cloud, self-hosted models, on-premises infrastructure, and hybrid arrangements—including systems built to meet defined sovereignty requirements.
Separate decisions
Application hosting and model hosting are separate choices. An application in your private cloud can call a managed model API. A self-hosted model can run in cloud infrastructure or on premises. A hybrid arrangement can keep selected processing local while using external services for permitted tasks.
The choices affect capability, cost, latency, and operating responsibility.
- Managed model APIs
- Provide model capability without running the model infrastructure yourself. Usage charges, service limits, provider changes, and data-handling terms remain part of the decision.
- Self-hosted models
- Give you more responsibility and control over model serving and updates. They also require suitable compute capacity, security, maintenance, and technical support.
- On-premises arrangements
- Can support local processing and connectivity requirements, with hardware capacity, resilience, and maintenance planned explicitly.
- Hybrid arrangements
- Divide work between environments, adding flexibility alongside integration and operational complexity.
Make control and ownership explicit
For the proposed design, we establish where inference and supporting services run; what information leaves each environment; who controls access, keys, and administration; and how prompts, embeddings used for search, outputs, logs, and backups are handled.
Keep retrieval within the right access boundaries: connect it to identity and source permissions, maintain permission changes, and test that users cannot obtain information outside their role.
We also agree who maintains the system, who can change it, and what moving to another provider or environment would involve. Model substitutions are evaluated against the application’s tasks.
Where sovereignty is a requirement, we translate it into specific jurisdiction, control, access, and continuity requirements. Hosting location or self-hosting alone does not establish that those requirements are met.
More Examples of Working Applications
Working applications
Connect the sources, the task, and the review.
Use the business’s existing systems and approval processes to put the right information in front of the person doing the work.
Handle a customer return request.
A service coordinator needs to respond to a customer whose purchase record is incomplete.
- Sources and workflow
- Retrieve the current return procedure and reviewed resolved cases, then check live order and product records through authorized connections. Apply explicit eligibility rules and prepare a response with unresolved questions highlighted. Past cases provide context without overriding current policy.
- Permitted actions and review
- Read authorized sources and prepare response and case-note drafts. Show the policy version, order fields and lookup time, and missing evidence. The coordinator corrects the draft or escalates to the service owner; refunds, policy exceptions, and customer communications remain with staff.
- Evaluation
- Test superseded policies, exact product identifiers, incomplete records, cross-customer access attempts, and conflicting guidance. Assess source support, correct escalation, correction effort, and handling time against the existing process.
An assistant with a defined sequence of lookups is a practical starting point. Adaptive tool selection earns its place when varied enquiries require it.
Turn a quotation into a purchase requisition.
An operations coordinator prepares a request through the existing purchasing process.
- Sources and workflow
- Extract quotation details with page references, retrieve approved purchasing procedures, and check current supplier, cost-centre, and approval information. Validate fields and calculations, resolve ambiguities, and prepare the requisition.
- Permitted actions and review
- A pilot produces drafts. A production integration may create a requisition after submission authorization and read its status. Show extracted values beside the quotation. The designated approver retains the purchasing decision; supplier payment details, purchase orders, and payments remain outside the application’s permission.
- Evaluation
- Test poor scans, currencies and decimal errors, missing terms, duplicate requests, revoked permissions, and interrupted writes. Confirm that the destination record matches the authorized draft and follows the correct approval route.
A timeout can leave the result unknown. Check the receiving system before retrying, and keep draft prepared, requisition created, awaiting approval, and approval confirmed visibly distinct.
Follow the purchasing workflowFor either application, budget includes source preparation, integration, reviewer time, and ongoing ownership alongside model usage. Begin with the sources and connections the task requires.
Where to start
Start with the decision or implementation in front of you.
Assessment and implementation are independent starting points. Bring a representative task, the systems and information involved, and the constraints the work must respect.
Assess the opportunity and choose an approach
Examine a defined task, its users, information, constraints, and current performance. Assess readiness, feasibility, and intended value to decide what is worth implementing.
- A recommendation with business requirements and acceptance criteria
- An information and integration map, architecture options, and cost assumptions
- An implementation plan matched to the task and available capacity
Build, connect, or improve the application
Begin with an established use case, a focused pilot, an existing application, or a release that needs engineering work. Implement the required connections and controls, test the workflow, and prepare user guidance and operational handover. Agree continuing support around the system and responsibilities involved.
- A working application or pilot meets the agreed task and scope
- Evaluation covers source support, access controls, authorized actions, and exceptions
- Release criteria, user guidance, and operating responsibilities are ready
The wider context
Include the environment the application depends on.
A production application also needs a dependable operating environment. Where hosting, release, or monitoring changes are required, bring them into the application scope.
Cloud operationFind an individual specialist for a defined AI, data, integration, or application role. Your team sets priorities and directs the specialist’s work.
Discuss your staffing needs
Start with the challenge
Which task could AI help your team complete?
Tell us what is happening today, what you want AI to make possible, and the constraints the work needs to respect.