At Data Friendly Space, we're acutely aware of the challenges faced by our humanitarian partners. These include an overwhelming influx of data from various sources, the need for resources to access, extract, and analyze data for efficient and quick decision-making (especially during emergencies), the ever-growing gap between humanitarian needs and financing, and the lack of easy-to-use data tools for staff working in often low-resource settings.
With these challenges in mind, we first conceptualized the GANNET Virtual Assistant in 2023, before the explosion of Large Language Models and Generative AI in our daily lives. The idea was to create a chatbot that humanitarian and development professionals could use to instantly access and extract information from a wide variety of resources such as news, social media, and humanitarian information platforms, without the need for resource-intensive labeling processes.
In November 2023, we were awarded as a winner of the 2023 Amazon Web Services (AWS) IMAGINE Grant. This enabled us to start the development process of the Virtual Assistant, an end-to-end generative AI-powered product designed to assess, extract, and visualize life-saving information to boost the timeliness, effectiveness, and accuracy of aid responses worldwide.
The Need for a Robust Development Methodology
In the rapidly evolving landscape of humanitarian aid and AI technology, the integration of generative AI presents both unprecedented opportunities and unique challenges, especially when the outcomes directly impact the lives of vulnerable communities. To effectively harness the potential of these advanced tools while ensuring they meet the complex needs of humanitarian contexts and maintain trust, robust design and development methodologies are crucial.
To ensure the Virtual Assistant's success and avoid falling into the "build trap," we have adopted the Dual-Agile product development approach. This methodology for user-centered product development, when applied to generative AI projects in humanitarian aid, offers a powerful framework for creating solutions that are both innovative and deeply attuned to the needs of affected populations.
The Dual-Agile Approach
The Dual-Agile product development methodology provides a dynamic framework for bringing generative AI solutions to life. By running parallel tracks of discovery (focused on product-user fit) and delivery (centered on product development), we can rapidly iterate on the Virtual Assistant's features while continuously validating their relevance, trustworthiness, and effectiveness in different scenarios. This approach keeps us focused on user needs and challenges while developing the Virtual Assistant's features.
We follow these two tracks simultaneously, allowing for continuous feedback and adaptation:
Discovery Track: Uses Lean Startup methodology
Focus: Product-user fit
Activities: User discovery, problem validation, solution testing
Delivery Track: Uses Scrum or other Agile development methods
Focus: Product development and iteration
Activities: Sprint planning, development, testing, and deployment
Uncertainty Management: Ideal for new technologies where market needs are unclear. Generative AI is relatively new to both private and humanitarian spaces. While most users are very interested, the needs and trust towards the technology are still evolving.
Rapid Learning: Fast iterations allow a quick understanding of complex AI capabilities and user needs.
Flexibility: Adapts to changing requirements common in emerging tech fields, especially when we hear about new features in LLMs and new methodologies with higher capabilities every month.
Risk Mitigation: Early and continuous validation reduces the risk of building the wrong product and addressing incorrect needs.
Efficient Resource Use: Focus on validated features prevents wasted effort on unnecessary development.
Innovation Fostering: Encourages experimentation while maintaining product development momentum.
User-Centric Design: Ensures AI products are developed with real user needs in mind.
Ethical Considerations: Allows for continuous evaluation of AI ethics and societal impact.
Scalability: Supports gradual scaling as the AI product matures and gains traction.
Competitive Edge: Enables faster time-to-users while ensuring product-user fit.
How We Adopt the Dual-Agile Process for the Virtual Assistant's Success
Hypothesis-Driven Development: We start with assumptions and test them. Our first hypothesis comes from our experience with projects where humanitarian staff spend most of their time searching and extracting data from vast amounts of sources to gain insights into current crises. We conduct user interviews and user-needs research to understand and validate these needs, which helps us specify and prioritize the Virtual Assistant features.
Rapid Prototyping: We focus on the quick creation of Minimum Viable Products (MVPs). Instead of developing all envisioned features at once, we selected the most needed and significant features to test in our MVP. These selected features come from validated user interviews, user research, and field experience. We keep the remaining features in the product backlog to re-prioritize and develop after continuous feedback.
User Feedback Loops: We maintain regular interaction with users for validation. After releasing our first MVP in May 2024, we welcomed beta users to validate and provide feedback on the Virtual Assistant. We follow up closely with regular monthly feedback sessions, surveys, and feedback mechanisms built into the Virtual Assistant platform. DFS conducts individual interviews with beta users to focus on user-specific needs and challenges. We're also working on specialized surveys and indicators such as new user acquisition, number of prompts generated, and user retention to help us understand the Virtual Assistant's current and future use, as well as new requirements and challenges.
Iterative Development: We constantly refine based on new insights. User feedback helps us understand how to adapt current features to address challenges and what to prioritize in upcoming feature development. For instance, for explainability, we initially provided the first 15 resources used to answer a user's question. However, most users found it difficult to validate the answer by checking each resource individually. In response, we're now working on providing in-answer citations, which will empower users to cross-check and understand more easily where the information comes from.
Cross-Functional Teams: We foster collaboration between business, design, and tech teams. At Data Friendly Space, we continuously develop the Virtual Assistant with a cross-functional team. Our tech team, responsible for building the Virtual Assistant's architecture and ensuring continuous deployment, consists of senior AI engineers, back-end and front-end developers, and DevOps engineers. User-centered design and communications are handled by our design and communications team. Needs assessment and performance testing are conducted by our humanitarian analyst team. The overall product development and technology advancements are guided by our product lead and CTO. This approach allows us to build and design the Virtual Assistant continuously in a synchronized and holistic manner.
Discovery Phase- Concepts
Initial Wireframe
Current Version
The Dual-Agile approach is particularly valuable for generative AI products like the Virtual Assistant, where technology capabilities and user applications are rapidly evolving, especially within the humanitarian space's unique needs and constraints. It allows us to stay responsive to technological advancements while ensuring the product meets real-world needs and expectations.
If you would like to join to shape the future of the Virtual Assistant please register to be a user here: https://app.gannet.ai/sign-up
.jpg)
