Machine Learning consultant

3. Objectives 

The consultant will: 

• Apply unsupervised machine learning (clustering) to client and population data from [Country A] to produce a needs assessment grounded in empirically-derived population segments rather than assumed demographic categories. 

• Conduct a methodological assessment and attempt at applying Multilevel Regression and Post-stratification (MRP) techniques to estimate needs in areas with sparse or uneven data coverage, reporting both what the estimates show and the confidence that can reasonably be placed in them. 

• Conduct a data feasibility assessment across candidate programme datasets in [Country B] to identify the most suitable pilot for predictive modelling. 

• Produce a methodological guide for IRC Regional Measurement Advisers (RMAs) drawing on both workstreams, enabling them to assess data readiness, commission or conduct similar analyses, and communicate findings to programme and business development audiences in future contexts. 

• Build, validate, and document a first predictive model for the selected programme following recognised best practices for model development, robustness testing and responsible deployment. 

4. Scope of Work 

Workstream A: Needs Assessment Clustering – [Country A] 

• A1 Data Assessment and Preparation. Review available datasets — likely including client registration data, Multi-Sector Needs Assessment (MSNA) outputs, and programme monitoring data — and assess quality, completeness and suitability for clustering. Produce a brief data assessment note (2–3 pages) documenting sources used, limitations, and any assumptions made in preparation. Conduct necessary data cleaning and structuring. This note will serve as a formal sign-off point before analytical work proceeds. 

• A2 Clustering Analysis. Apply an appropriate clustering algorithm (eg K-medoids, hierarchical clustering, or similar) to identify distinct population segments. The choice of method should be justified and documented. In parallel, conduct a methodological assessment of whether MRP techniques can extend the reach of the analysis to data-sparse locations. The consultant should attempt an MRP model where data conditions permit, report the resulting estimates with explicit uncertainty intervals, and document clearly what the estimates show, the limitations of the approach given available auxiliary data, and what additional data collection would be required to improve reliability. The MRP component is understood to be exploratory; its value lies as much in surfacing the boundaries of current evidence as in producing point estimates. 

• A3 Needs Assessment Report. Produce a written needs assessment report interpreting the cluster outputs in programmatic terms. The report should translate statistical findings into actionable insights — describing who the population segments are, what their distinct needs appear to be, and what the implications are for programme design and targeting. Where MRP estimates are included, they should be clearly labelled with appropriate caveats. The IRC's MEAL and programme staff will contribute to contextual interpretation and the final write-up; the consultant leads the analytical narrative. 

Workstream B: Predictive Modelling Pilot – [Country B] 

• B1 Data Feasibility Assessment. Review candidate programme datasets — likely spanning two to three programme areas (potentially including economic recovery and development, education, or protection) — and assess their suitability for predictive modelling. Suitability criteria include: availability of historical outcome data, sample size, data completeness, and ethical considerations around the use of predictions in that programmatic context. Produce a brief feasibility note recommending which programme to proceed with and why, agreed with the IRC before model development begins. 

• B2 Model Development and Validation. Following agreement with the IRC on the selected programme, build a predictive model using historical programme data. The model should predict a clearly defined outcome (eg programme completion, dropout risk, or a specific client-level result) using data available at or shortly after client intake. 

• Model development must follow recognised best practices throughout, including: 

• A documented train/test split, or cross-validation where sample size constrains a held-out test set, with performance metrics reported on data the model was not trained on. 

• Robustness checks across key demographic subgroups — at minimum sex and age, and where data permits displacement status and other contextually relevant characteristics — to identify differential performance and potential sources of bias. 

• Feature importance analysis and sufficient documentation of model internals to allow non-specialist reviewers to understand what the model is and is not doing. 

• Explicit documentation of the conditions under which the model should not be used, or should be retrained — including population shifts, changes in programme design, or data quality deterioration. 

• All analytical code written in R or Python, shared with the IRC in a commented and fully reproducible format, with a clear README. 

• B3 Operational Tool and Documentation. Translate the validated model into a lightweight operational tool — a scoring widget, dashboard integration, or structured output — that programme staff can use without specialist data science knowledge. Produce accompanying documentation covering: how the tool works, how to interpret its outputs, its known limitations, and recommended protocols for human oversight and override. The tool should be deployable within the IRC's existing Databricks and Power BI infrastructure. 

Workstream C: Methodological guide for regional measurement advisers 

• C1 RMA Methodology Guide Drawing on the experience and learning from both workstreams, produce a practical methodological guide (approximately 8–10 pages) intended for IRC Regional Measurement Advisers. The guide should be written for an audience that is quantitatively literate but not specialist data scientists. It should cover: how to assess whether available data is suitable for clustering or predictive modelling; rules of thumb for method selection; key questions to ask when commissioning or reviewing this type of work; how to sense-check analytical outputs; and how to present findings to programme and business development audiences. The guide should be grounded in the specific experience of this consultancy rather than generic methodology, and should be reviewed by at least one RMA before finalisation. 

5. Deliverables and Timeline 

Note: Workstreams A and B run in parallel. The timeline assumes data access is provided promptly at inception and that the data assessment sign-off (Workstream A) and feasibility note agreement (Workstream B) proceed without significant delay. Delays in data access or sign-off will affect the schedule accordingly. 

6. Ethical Considerations and Data Protection 

The IRC takes data protection seriously and operates under its organisation-wide data protection policy. The consultant will be required to sign a data sharing agreement prior to accessing any client-level data. All data provided will be de-identified. The consultant must not share, publish or retain IRC data beyond the terms of this agreement. 

Predictive models built under this ToR are intended solely to support the allocation of additional resources and tailored support — not to exclude clients from services or make automated decisions affecting individual cases. The robustness checks and subgroup performance analysis described in section 4B are a mandatory component of the work, not optional enhancements. The consultant is expected to engage with these ethical parameters from the feasibility assessment stage onward, and to reflect them explicitly in the tool documentation and any outputs shared with IRC programme staff. 

7. Required Qualifications 

• Advanced degree (Master's or PhD) in economics, data science, statistics, computer science, or a related quantitative field, or equivalent professional experience. 

• Demonstrated experience applying clustering and predictive modelling techniques to real-world datasets, with examples of work shared at application. 

• Proficiency in R or Python (required); experience with Databricks is an advantage. 

• Proven experience working with messy, incomplete, or administratively-generated data — not only clean research datasets. 

• Experience with MRP or similar small-area estimation techniques is desirable; candidates should indicate clearly their level of familiarity with this method. 

• Familiarity with the humanitarian or international development sector, ideally including direct experience working within an NGO or research institution operating in crisis contexts. 

• Strong technical writing skills and the ability to translate analytical findings for non-specialist audiences. 

8. Management and Reporting 

The consultant will report to [Philip Blue / Regional Measurement Adviser, Latin America] and work in close coordination with the IRC's Measurement Unit and relevant country programme MEAL staff. 

An inception call will be held within the first week to align on data access, workplan, and communication rhythms. Brief fortnightly check-ins are expected throughout. Two formal review points are built into the timeline: sign-off on the Workstream A data assessment note, and agreement on the Workstream B feasibility recommendation, both before substantive analytical work proceeds. The Workstream C methodology guide will be reviewed by at least one Regional Measurement Adviser in addition to the primary supervisor prior to finalisation. 

9. Submission Requirements 

• CV highlighting relevant data science experience. 

• A short expression of interest (maximum one page) describing your approach to one of the two workstreams. 

• Two examples of prior analytical work — published outputs, code repositories, or technical reports. 

• Daily rate in USD. 

• Names and contact details of two referees from previous clients or employers.