Selected implementations and technical evidence
A non-exhaustive selection of business systems, workflow implementations and technical evidence. Each example connects an operational problem with the engineering and controls needed to address it.
Across these projects, we combine AI-native systems, AI agents and AI-assisted workflows, shaped by the realities of each business.
Credit applications and approvals
Built a credit operating system connecting applications, supporting evidence, credit assessment and approvals.
Supplier invoices and payment approvals
Built accounts-payable workflows that check invoices, handle discrepancies and route approvals, keeping financial controls and supporting evidence within the process.
Coordinating complex supply chains
Built multi-level supply-chain orchestration that coordinates decisions across organisational levels and resolves competing priorities.
Cybersecurity across business operations
Built systems connecting security monitoring, investigation and coordinated response, with attention to business disruption and who has authority to act.
Intelligence and automation at the edge
Developed edge-computing applications for rural telecom networks, smart cities, healthcare and security, including automation close to where the work happens.
Building payment ecosystems
Built payment ecosystems that bring businesses, financial institutions and technology together to deliver payment services.
Common questions
Does every project use AI in the same way?
No. AI can assist a person, support a decision or carry out defined work through agents. The role and authority depend on the task.
Concrete workflows. Reviewable engineering.
These are non-exhaustive examples of the systems and control patterns we build. They show workflow behaviour and engineering boundaries across selected implementations.
Payment outcome verification: requested instructions, observed records and reviewable evidence.
We built API-based comparison of payment requests and outcome records. The workflow classifies matched, mismatched and awaiting-confirmation outcomes, preserves history when confirmations arrive late, rejects conflicting resubmissions and avoids duplicate history on repeated checks. Evidence and audit exports support review within tenant boundaries. The application verifies records; the relevant payment systems retain responsibility for fund movement.
Connected credit cases: intake, evidence, assessment and distinct approval responsibilities.
We connected intake, documents, extracted facts and assessment within a persisted credit case. The workflow separates evidence, policy, approval and execution authority, so assembling or assessing a case does not silently authorise a lending action. This reusable control pattern keeps the business process and its decision responsibilities explicit.
Accounts-payable workflow: extraction, validation, exceptions and controlled approval handoff.
We connected invoice extraction and validation with exception handling and approval routing. Server-side authentication, role-restricted approved-invoice exports, tenant boundaries and audit records keep access and review responsibilities alongside the workflow. Approval handoff remains distinct from authority to release a payment.
Frameworks, voice tooling, protocols and runtime infrastructure
Frameworks
FreshCtx documents protected-action patterns for Agno tool hooks, LangGraph action nodes and checkpoint resume, OpenAI Agents SDK custom function tools, and Google ADK before-tool callbacks.
Voice tooling
ElevenLabs Python ClientTools handlers provide a documented boundary for revalidating declared evidence before a configured client tool acts.
Protocols
Documented patterns cover MCP tools/call and receiving-agent A2A execution. The application retains responsibility for authentication, authority, retries and transaction behaviour.
Runtime infrastructure
The public Prudenze–OpenShell qualification report records bounded allow, block, escalation, stale-evidence and outcome-mismatch scenarios. It is implementation evidence for those scenarios, rather than a vendor certification.