Six practice areas. One method.
Each service is an integrated workflow, not a sales blurb. Engagements are scoped from the methodological problem backwards: what evidence is needed, what the data actually allows, and where AI tooling earns its place, whether that's a full evaluation commission or lightweight evaluative support embedded in a programme team's workflow.
Evaluation design
Right the way through the inception process. Methodologically rigorous, structurally sound, and built faster than the traditional cycle allows.
Inception, evaluation matrices, sampling strategy and mixed-methods frameworks grounded in OECD-DAC criteria. LLM-assisted literature scans accelerate document review across hundreds of programme reports without losing the analytical thread.
Instrument drafting (interview guides, survey schedules, observation protocols) is supported by structured prompting workflows, then reviewed and refined by hand. The output is an evaluation design that holds up to peer review and a workplan that fits real timelines.
- —Inception report
- —Evaluation matrix
- —Mixed-methods workplan
- —Instrument suite
Quantitative analysis
Reproducible survey pipelines and analysis notebooks, built once and re-run on demand as new data arrives.
End-to-end processing of household, beneficiary and key-informant surveys: ingestion from Kobo, ODK or CommCare CAPI; automated cleaning and validation; outlier detection; longitudinal merging; weighted analysis. Standard BHA indicators (FCS, HHS, rCSI) and Sphere WASH benchmarks are computed by default where the dataset includes them.
Quasi-experimental approaches (DiD, PSM, regression discontinuity) where the data supports them, and honest communication of uncertainty where it doesn't. Outputs are reproducible notebooks and lightweight dashboards, not a black-box deliverable.
- —Reproducible Python pipelines
- —Cleaning & validation reports
- —Statistical analysis & dashboards
- —Quasi-experimental designs
Qualitative analysis
LLM-supported thematic coding of interviews and FGDs, with human-in-the-loop validation and a fully traceable evidence chain.
Codebook development is iterative and human-led. Coding is then accelerated by language models, with every coded segment linked back to its source utterance. No orphaned claims, no synthesised quotes, no hallucinated themes.
A sample of every coding pass is re-coded by hand for consistency, and inter-coder agreement is reported transparently. The result: qualitative depth at a pace that matches programme decision cycles, without sacrificing methodological credibility.
The same approach scales to portfolio-level evidence synthesis: lessons drawn across multiple evaluations, country programmes, or thematic portfolios, identifying what works, what doesn't, and where the evidence gaps sit. Source-traced throughout, and flagged as preliminary until validated.
- —Codebook & coding frame
- —Coded transcripts with provenance
- —Inter-coder consistency report
- —Thematic synthesis
- —Portfolio evidence syntheses
Reporting and quality assurance
Structured drafting workflows that compress reporting timelines while raising the bar on internal consistency.
Structured drafting moves from evidence base to outline to prose in clearly separated stages. Automated checks verify citation integrity, internal consistency between findings and recommendations, and alignment with the evaluation matrix.
Plain-language summaries and rapid translation review across working languages are part of the standard workflow. Human editing remains essential: the AI handles the mechanical work so the evaluator can focus on judgement.
Validation workshop slide decks, dissemination briefs, and two-page summaries, designed in programme brand systems or Meridian's own, with every figure drawn directly from the underlying analysis. Branded knowledge products produced in hours, not days, with an embedded QA layer that ensures no number on any slide is unverifiable.
- —Evaluation reports & annexes
- —Citation & consistency QA
- —Plain-language summaries
- —Translation review
- —Workshop presentations & briefs
- —Branded knowledge products
Custom MEL tools & dashboards
Bespoke tooling for implementing partners: small, focused, and built to be handed over.
Indicator dashboards, beneficiary feedback pipelines, document Q&A systems over programme archives, and lightweight evaluation knowledge bases. Anyone can upload documents to a language model. The difference is in the prompt architecture: structured system prompts that encode evaluation frameworks, programme terminology, donor reporting conventions, and the specific analytical lens the evaluator needs. Meridian builds these retrieval layers so the model returns source-attributed, methodologically grounded answers, not generic summaries of whatever it found. Built with open components where possible, deployed where the data needs to live, and documented for handover.
These are not platform sales. The brief is always a specific friction point in a partner's MEL workflow, and the deliverable is a tool that solves it, not a subscription.
- —Indicator dashboards
- —Document Q&A systems
- —Evaluation knowledge bases
- —Beneficiary feedback pipelines
Fieldwork support & data operations
The operational layer: transcript processing, survey tracking, and respondent management during live data collection windows.
AI-assisted transcript processing from recorded interviews: accent correction informed by programme context, responses mapped to evaluation sub-questions, structured templates filled with near-verbatim responses, and substantive findings flagged as they emerge. Train the model on your full transcript set and it identifies convergence, divergence, and outlier perspectives across stakeholders, with source attribution to specific respondents.
Real-time survey progress tracking from CAPI exports: quota monitoring disaggregated by site, grade, gender, and survey instrument. Automated error detection including mislabeled entries and submissions at non-sampled sites. Daily status reports during fieldwork windows so team leads know exactly where the gaps are before the field team moves on. The same pipeline handles logframe indicator tracking between collection rounds: progress against target by indicator, on-track and at-risk areas summarised, and trends across reporting periods produced from the data rather than assembled by hand from spreadsheets.
Beneficiary feedback and complaints data processed and categorised at scale, with trend identification and reporting-ready summaries, so that feedback mechanisms produce something a programme team can act on rather than a backlog nobody reads.
Client comment management on draft deliverables: comments categorised by type (factual correction, methodological challenge, editorial preference, scope change), responses drafted, and those requiring substantive revision flagged for evaluator attention. Multi-language transcript correction that goes beyond generic speech-to-text: the model corrects accent-driven errors using programme context, distinguishing organisation names, local place names, and sector terminology that standard transcription tools consistently misrender.
- —Cleaned, EQ-mapped KII/FGD transcripts
- —Cross-stakeholder analysis summaries
- —Daily sampling progress trackers
- —Respondent mobilisation matrices
- —Comment tracking & response logs
- —Context-corrected multilingual transcripts
- —Logframe indicator tracking
- —Feedback & complaints analysis