Evidence for Development Finance.

Development Finance runs on evidence, and most of it sits buried in thousands of PDFs, spreadsheets and old reports. Destillo is a methods firm. We design the method, build the pipeline and deliver structured, benchmarked findings across your whole portfolio. Your evaluators keep the judgment.

illustrative sector portfolio, 61 operations reading
24,812pages read
800guarantee-related references
300climate-related references
ENESFRPT
page read in full guarantee-related passage climate-related passage Nothing sampled. Nothing rated by the system.
3.7xmore evidence found than a basic LLM setup, on the same document
90%less manual analysis time in production deployment
100%of pages covered. Nothing sampled, nothing skipped
2,000,000+pages read in production
Up to 99.7%recall across benchmarked engagements

Institutions our team has delivered for

Where we have presented our methods

Accuracy

We tested our pipeline against MAI, the World Bank Group's internal RAG system, on the same project appraisal document.

MAIWorld Bank Group internal RAG system
9 found 24 missed
27%recall
Destillo pipelinehybrid retrieval, then classification
33 found
99.3%recall
activities found and verified activities missed

33 verified resilient infrastructure activities in the document. MAI found 9 and missed 24. The misses clustered in N-1 reliability criteria, SCADA redundancy, disaster risk insurance, sedimentation studies and equipment lifecycle.

Note: This comparison was done in 2025, and the technology keeps changing.

Applications

Some of the work we have already done with evaluation, results and outcomes teams.

Validating self-evaluation

Check what a completion report claims against what the record actually shows. Divergences go to the validator.

Quality of M&E systems

See whether indicators have baselines, named data sources, and the endline evidence that was promised.

Portfolio and sector synthesis

Ask one question across hundreds of operations. Get an answer for each, traced to the paragraph.

Outcomes-based contracting

Trace outcome metrics through contract, baseline and provider records. Check payment triggers against independent evidence.

Outcomes

Lessons and precedent

Find what the institution already concluded on this design, sector and country, before the next appraisal.

Multilingual archives

Read Spanish, French and Portuguese filings alongside English, including decades of scans. Citations point at the original.

Keeping verification at the centre

A typical architecture looks like this. A two-way road with a roundabout at every stage. The machine finds and drafts, a person checks, and any stage can go back. The specifics change with the task. The shape does not.

    Moving forward Looping back Revisit earlier decisions

    These are task-specific systems, not a product you log into. Each pipeline is built around one institution's questions, rating criteria and document conventions. We specify those questions with your team, test the pipeline against work the office has already completed, and put it into production in your environment. We run it. Your evaluators get the final, clean answers, each with its citations.

    Ask a portfolio a question

    One question, applied to a whole sector portfolio instead of the four operations someone had time to read. Below is one round of an engagement. Our team runs it with yours, and every round feeds the next.

    1. 1Agree the questions and rating criteria with your team
    2. 2We run them across every page of the portfolio
    3. 3Your evaluators review the findings with us
    4. 4Refine the questions and run the next round

    Example questions from a past round. Open one to see what comes back.

    61operations
    24,812pages read in full
    4filing languages
    3,140scanned pages, no text layer
    Open an example question to see one round of findings.
    example round

    Every page is read, including the scanned annexes and the Spanish and French filings. Findings come back attached to the paragraph, table cell or survey field they came from.

    A run at 50,000 pages works the same way as a run at 250. Cost and time scale with the corpus. The method does not change.

    An illustration, not a product. Synthetic portfolio constructed for illustration. No client material is shown. Vocabulary generalised across MDB and DFI practice.

    Case studies

    Our team has spent four years building machine learning and AI systems inside multilateral development banks, including the World Bank Group and the Asian Development Bank. Five of those engagements are below, each run on real institutional material at portfolio scale.

    Asian Development Bank · IEDM&E quality

    Quality of M&E systems across small island developing states

    Assessing how well monitoring and evaluation systems were designed and used across operations in small island developing states, where thin statistical capacity and small sample sizes make results reporting hardest to verify.

    Twenty evaluation questions were applied to every page of the corpus. Baseline presence, indicator measurability, data source specification, endline collection and sixteen others. The result is a comparable assessment per project rather than a narrative impression across the group.

    Projects225
    Documents750
    Pages50,000
    Questions20
    World Bank Group · IEGEvaluation

    World Bank Group Guarantees

    An evaluation of the guarantee portfolio required consistent extraction from a document set spanning instruments, regions and two decades of changing templates. Twenty-seven distinct questions were applied across all 25,000 pages, with no sampling.

    The methodology was benchmarked head to head against manual review of the same documents. The evaluation team received portfolio-wide coverage, plus exact figures on where the pipeline agreed with reviewers and where it did not. That reporting is what makes an accuracy claim usable rather than promotional.

    Documents241
    Pages25,000
    Questions27
    SampledNone
    Asian Development BankClassification

    Finding Digital Projects

    Digital work is rarely labelled as such. It sits inside transport, health and public sector operations as a component, so the portfolio cannot be identified from project titles or sector codes alone.

    All 5,500 documents were classified for whether the project belonged in the digital portfolio, with the decision and its basis recorded per document, so the inclusion boundary could be reviewed and defended rather than assumed.

    Documents classified5,500
    Coverage100%
    Decision recordPer document
    World Bank GroupReplicability

    Replications across MDBs

    A study of replicability across development finance portfolios: which project designs recur, which correlate with results, and which are specific to their context.

    Ten thousand projects were assembled from more than ten data sources across five institutions, including EBRD. Embedding-based similarity identified comparable operations across institutional boundaries, and surfaced correlations that keyword matching on five different taxonomies could not.

    Projects10,000
    Data sources10+
    Institutions5, incl. EBRD
    World Bank Group · GFDRRCompliance · Multilingual

    Automated Building Code Compliance Review

    Building documentation had to be assessed against a structural safety and compliance checklist, across a document set filed in several languages. More than fifty structural questions were built into the pipeline, each returning an assessment traceable to the drawing, specification or report clause it came from.

    Translation was built into the pipeline rather than bolted on. Non-English submissions were assessed against the same checklist as English ones, and the citation kept pointing at the original text. This is the engagement the multilingual capability came out of.

    Structural questions50+
    TranslationIn-pipeline
    CitationsSource language

    Describes work delivered by the Destillo team. Not an endorsement by any institution. No confidential material is reproduced here.

    How we work

    We work as consultants alongside your team. Start small, prove it on your own material, then run it every year.

    Pilot

    A fixed scope on one live task, such as a set of validations. We specify the questions with your team, build the pipeline, and benchmark it against work your office has already completed.

    Weeks, fixed fee

    Production

    The pipeline moves into your environment and we run it on new work, against the accuracy and time-saving targets agreed in the pilot. Your evaluators get final answers with citations.

    In your tenancy

    Annual engagement

    Recurring validations, syntheses and data collection run on one maintained system. New questions are added as your work programme changes.

    Recurring

    Privacy and data handling

    The questions an AI governance review will ask, answered before the first meeting.

    GDPR

    Personal data is handled under GDPR, with processing agreements and data subject rights set in the engagement terms.

    Encrypted at rest and in transit

    Documents and outputs are encrypted while stored and while moving between systems.

    Your IT and governance frameworks

    We work within each institution's IT, information security and AI governance requirements, and complete your review before any data moves.

    Deploys in your tenancy

    The pipeline runs inside your own cloud environment where the engagement requires it.

    Zero retention

    Documents are purged after processing. Nothing is kept beyond engagement scope, and nothing trains a model.

    Full audit trail

    Every retrieved passage, proposed finding and human override is logged and reconstructable.

    Outputs map to IRIS+, HIPSO, OPIM, IFC Performance Standards, SASB and SDG reporting where required.

    Start with a pilot today.

    Pick a document set and tell us the questions that matter. Book a free 30 minute call to get started.