01
Workflow automation
Removing manual steps from repetitive processes — data entry, status updates, routine correspondence — while keeping approvals with the people accountable for them.
AI Solutions
AI that improves real workflows, under real oversight.
We apply AI to specific, well-understood tasks inside existing processes: reading and sorting documents, answering questions from approved sources, routing requests, and preparing work for people to review. Every deployment is evaluated, logged, and reversible.
What we build
Seven kinds of AI work we deliver regularly.
Most are built on commercially available or open models, configured and constrained for your data and rules. We do not train foundation models; we make them useful and safe in context.
02
Document processing
Classifying, extracting, and validating information from forms, applications, invoices, reports, and scanned records, with confidence scores and review queues for uncertain cases.
03
Intelligent search and knowledge workflows
Question answering over approved internal documents, policies, and records, with every answer citing its source so staff can verify before acting.
04
Assistant and triage systems
First-line handling of inquiries by email, web, or messaging channels: understanding the request, gathering what is needed, and routing it to the right team with a summary.
05
Review and summarisation
Condensing long documents, case histories, meeting records, or submissions into structured summaries that highlight what a reviewer needs to check.
06
Supervised decision support
Presenting relevant evidence, precedents, and rule checks alongside a recommendation — the decision stays with a person, and the reasoning is recorded.
07
Integration with existing systems
Connecting AI components to your APIs, databases, records systems, and messaging channels so outputs land where work happens, not in a separate tool nobody opens.
How a supervised workflow looks
AI prepares the work. People decide.
A conceptual view of a typical document intake workflow. The highlighted steps are where staff review and approve before anything is recorded or sent.
- IntakeDocuments arrive by portal, email, or scan. Each is logged with a reference and timestamp.
- Classify & extractThe AI component identifies document type and extracts key fields, with a confidence score per field.
- Rule checksDeterministic checks run against business rules: completeness, format, duplicates, eligibility criteria.
- Officer reviewA staff member sees the extracted data, source document, flags, and confidence. They correct, approve, or reject.
- Record & notifyOnly after approval is the records system updated and the applicant notified. Every step is in the audit log.
Automated step, logged
Human review and approval
Safeguards
What makes AI acceptable in institutional settings.
These are not optional extras. They are part of how we design, estimate, and test every AI deployment, and they are what oversight functions ask about first.
-
01
Human oversight
Each workflow defines where a person reviews, approves, or can override. Consequential actions — updating an official record, sending a decision, making a payment — require explicit approval unless the client's governance authorises otherwise.
-
02
Evaluation and testing
Before launch we build an evaluation set from real, anonymised examples and measure accuracy, failure modes, and bias indicators. Results are shared in plain language, and the same tests run after every change.
-
03
Security and privacy
Data minimisation, encryption in transit and at rest, scoped access, and clear rules about which data may be sent to which model or provider. Where required, we design for in-country hosting or on-premises deployment.
-
04
Traceability
Every automated output is logged with its inputs, sources, model version, and the person who acted on it. Answers from knowledge systems cite their sources. Auditors can reconstruct what happened and why.
-
05
Reliability and fallback
Systems degrade gracefully: if a model is unavailable or uncertain, work routes to people rather than stopping or guessing. Monitoring alerts on error rates, latency, and drift in output quality.
Readiness and rollout
Start small, measure, then extend.
Most organisations are not ready to automate everything at once — and should not try. We recommend a readiness assessment followed by a phased rollout with a clear decision point at each stage.
Readiness assessment
Review of candidate workflows, data quality, systems access, governance requirements, and staff capacity. Output: prioritised use cases and a go/no-go view.
Pilot design
One workflow, clearly bounded. Success criteria, review points, evaluation set, and privacy assessment agreed in writing.
Controlled pilot
Running alongside the existing process with a small group of staff. Every output reviewed; metrics and feedback collected weekly.
Evaluation and decision
Results compared against criteria. Decide to extend, adjust, or stop. Findings documented for governance and procurement.
Scaled deployment
Broader rollout with training, monitoring, support arrangements, and a schedule for re-evaluation as data and rules change.
Technology approach
Model-agnostic, integration-first.
Choice of models
We select models based on the task, data sensitivity, cost, and hosting constraints — commercial APIs, open-weight models deployed privately, or a combination. Clients are not locked to a single provider.
Data residency
Where policy or law requires it, we design for data to remain in Rwanda or within the client's infrastructure, including self-hosted model deployment and strict controls on what leaves the environment.
Built on your systems
AI components are delivered as services that integrate with existing APIs, databases, identity providers, and messaging channels, with the same security standards as the rest of your estate.
Maintainable by design
Prompts, rules, evaluation sets, and configuration are versioned and documented. Your team can understand, adjust, and re-test the system without us.
Common questions
Questions we are asked about AI projects.
Do we need clean, structured data before starting?
No, but you need to know where your data is and who owns it. Many useful AI workflows start from unstructured documents and email. Part of discovery is assessing data quality honestly and designing around its limitations.
Can AI systems run without sending data to overseas providers?
Yes. Open-weight models can be deployed on infrastructure you control, in Rwanda or in your own data centre. This affects cost and capability, and we will set out the trade-offs clearly during discovery.
How do you handle Kinyarwanda and other local languages?
Model performance varies by language. We test on real examples in the languages your service uses, design review points accordingly, and are transparent about where human handling remains necessary.
What happens if the AI makes a mistake?
The workflow is designed so that mistakes are caught at review points before they have consequences, and every output is logged so errors can be traced and corrected. Error rates are measured continuously, and any increase triggers review.
Will staff lose their jobs?
Our projects are scoped to remove repetitive effort and backlogs, not to remove accountable people from decisions. In practice, staff time shifts towards review, exception handling, and service quality. We work with you on change management and training as part of delivery.
Start with an assessment
Find out which of your workflows are ready for AI — and which are not.
A readiness assessment gives you a prioritised list of use cases, the risks attached to each, and a realistic plan. It is the right first step for most institutions.