4. Development phase: Model development
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna.
Why should I do this?
An iterative process where a variety of models are derived, tested, and refined until an appropriate model that meets the desired criteria is created.
- Workflow activities Miro Board
Target users
- CTOs and technical leaders
- Data scientists and ML engineers
- Product managers
- Research organizations, think tanks
Checklists in play
- Checklist 1: Risk Identification & Mitigation (Q9, Q10, Q11, Q12)
- Checklist 3: Legal & Regulatory Compliance (Q9, Q10, Q11, Q12)
Questions to consider
Checklist 1 · Q9
What measures have you taken that will help users interpret model outcomes and understand its limits?
Checklist 1 · Q10
What measures have you taken to test the model for safety and robustness against adversarial inputs?
Checklist 1 · Q11
What privacy-preserving techniques have been applied to protect individuals from re-identification or inference risks
Checklist 1 · Q12
How are changes to the model documented, tracked, and approved throughout its lifecycle?
Checklist 3 · Q9
In case a third-party contractor (vendor) is being used for any stage of the design, development, deployment, or post-deployment processes, what measures are being taken to ensure such vendor’s compliance as your organization's agent? Have these legal obligations been clearly allocated and contractually documented?
Checklist 3 · Q10
If using open-source or foundation models, have licensing obligations been assessed?
Checklist 3 · Q11
Has internal accountability for legal and regulatory compliance, and monitoring been allocated at executive level, and into the broader product development process?
Checklist 3 · Q12
Is there a documented protocol for incident reporting and regulatory notification?
Investment types
Every investment project is unique
The application of the six steps will vary accordingly. To provide examples that align with your project, common characteristics of AgDev investments were researched and three ‘investment types’ were developed.
AgriConnect: a digital solutions investment
The team configured a RAG LLM connected to curated agricultural knowledge bases.
Internal testing was conducted using multilingual prompts and extension officer scenarios.
Participatory co-creation, feedback, and model refinement
- Are domain experts involved in validating recommendations?
- What is the process to flag confusing or culturally inappropriate outputs and prevent it from recurring?
- What are the checks in place to ensure the model being developed is aligned to the user needs defined during the design phase?
Lessons learned and challenges that arose
Testing revealed that the model occasionally generated highly confident but unsupported recommendations. Some agricultural terms were mistranslated due to insufficient local language training data.
Decisions taken and documentation maintained
The team introduced retrieval constraints requiring outputs to reference curated agricultural sources rather than relying solely on model memory. Human review checkpoints were added for high-risk recommendation categories. These safeguards were documented in model governance procedures, refinement policies and escalation matrices.
AgroThrive: a policy and advocacy investment
The team configured a RAG LLM and tested synthesis outputs using real policy and investment briefing scenarios.
Participatory co-creation, feedback, and model refinement
- Are policy analysts, economists and sector experts involved in validating synthesis quality?
- Could users challenge framing bias in outputs?
- What historical or systemic biases did user research highlight, and what is being done to actively test the model to mitigate it?
Lessons learned and challenges that arose
Internal testing revealed that the model often generated highly coherent narratives even when underlying reports contained conflicting recommendations. Users perceived these outputs as more credible because the contradictions had been smoothed out.
Decisions taken and documentation maintained
The team prohibited single answer synthesis for contested policy questions. Outputs were redesigned to surface competing perspectives and confidence levels explicitly. Governance decisions and testing findings were documented in evaluation reports and model refinement logs.
NGBT: a field research investment
The team configured a RAG GenAI assistant capable of answering farmer questions, generating educational explanations and comparing crop options.
Participatory co-creation, feedback, and model refinement
- Are domain experts involved in validating generated outputs?
- Can users challenge persuasive framing or biased messaging?
- How does the model perform on specific user subgroups identified as vulnerable or high-priority during the design phase?
Lessons learned and challenges that arose
Internal testing revealed that the model naturally adopted highly optimistic language patterns because most source materials emphasized benefits and innovation outcomes. Some outputs subtly framed existing farming approaches as less advanced or less desirable.
Decisions taken and documentation maintained
The team introduced balanced response constraints requiring outputs to acknowledge uncertainties, resource requirements and contextual suitability. Human reviewers were required to validate high-impact messaging scenarios. These controls were documented in model refinement policies and ethical communication standards.