Research project · Nairobi, Kenya
AI for verifying disaster alerts, not amplifying them.
FlowOps Africa is building an AI system to verify disaster alerts and reduce misleading information. This page describes a focused reliability experiment: a small prototype, structured evaluation, and practical safeguards for high-stakes AI outputs.
Lead
Ismael Kiprop, Founder & Lead AI Engineer
Organisation
FlowOps Africa · Nairobi, Kenya
Status
Controlled experiment · not a live warning service
Focus
AI reliability, evaluation, and deployment safeguards
The problem
Convincing AI outputs are not necessarily accurate
AI systems can generate fluent explanations and recommendations that look authoritative. In disaster response, unverified AI-generated information can cause confusion, undermine trust, or push people toward unsafe decisions.
This project investigates practical methods for evaluating AI reliability and reducing the chance that unsupported outputs are treated as verified information. The initial application is disaster information; the methods may also inform other high-stakes deployments.
Approach
Prototype first, evaluate carefully
The initial phase builds a small prototype and a structured evaluation process. The system will be tested against disaster-related scenarios chosen to surface ambiguity, conflict, and misleading cues.
What we measure
Evaluation criteria
The project examines whether AI-generated alerts meet a clear bar for responsible communication under incomplete or contested information.
Evidence support
Whether an AI-generated alert is factually supported by the available reports and environmental context.
Uncertainty expression
Whether the system appropriately communicates confidence, incomplete evidence, and open questions.
Verified vs unconfirmed
Whether verified information is clearly distinguished from rumours, hearsay, or unconfirmed reports.
Severity and location claims
Whether the system avoids unsupported claims about where an incident is and how severe it is.
Safeguards under test
Reducing unsupported confidence
- Evidence checks before an alert is treated as verified
- Uncertainty labeling on incomplete or contested outputs
- Source tracking for claims used in alert summaries
- Human review for high-impact or high-risk alerts
Intended outputs
What this project will produce
Working proof of concept
A small prototype that interprets disaster-related reports with explicit reliability controls.
Evaluation dataset
A structured set of scenarios designed to probe failure modes, not only happy-path cases.
Documented test results
Recorded successes and failures so others can inspect how the system behaves under pressure.
Practical safeguards
Guidance for handling uncertain, conflicting, or potentially harmful AI outputs.
Why FlowOps
Practical AI engineering from Nairobi
FlowOps Africa builds software, AI integrations, APIs, and automation for African businesses. That engineering foundation supports a limited, testable reliability prototype rather than a premature production warning system.
The approach is deliberate: define measurable criteria, test against realistic scenarios, record failures, and improve safeguards from evidence.
Why now
AI-generated information is getting easier to believe
As AI outputs become more fluent and more widely available, the cost of treating unsupported claims as verified rises. Establishing ways to test reliability and communicate uncertainty before deploying AI in consequential settings is therefore timely.
Designed with African operating contexts relevant to disaster information: incomplete connectivity, multilingual reporting, and communities that need clear uncertainty signals.
Evidence of work
Links for reviewers
Use these links in grant applications, technical feedback, or follow-up conversations about the prototype and evaluation plan.
Canonical page URL: https://www.flowops.africa/projects/disaster-alert-verification