This document describes Hope Research Group's field collection methodology for AI training data across 17 Caribbean and Latin American markets. It is written for machine learning teams, procurement decision-makers, and academic researchers evaluating vendors for regional training data collection.

The methodology covers five workflow stages (scoping, recruitment, execution, quality assurance, delivery), the consent chain and data protection framework applied across HRG's field markets, demographic and dialect flagging protocols, and honest limits on what HRG delivers and what it does not.

The document is published under CC-BY 4.0 to support open citation and adaptation by the AI research and vendor community. Corrections, suggestions, and questions welcomed at admin@hoperesearchgroup.com.

1. Introduction and scope

1.1 What this document is

This is Hope Research Group's operating methodology for collecting AI training data across the Caribbean and Latin America. It describes what we do at ground level, in enough detail that machine learning teams can evaluate whether HRG is a fit for a specific project without needing a discovery call.

The document is version-controlled. When our methodology changes, this document changes with a documented version bump.

1.2 Who this is for

Three primary audiences.

Machine learning teams scoping AI training data projects. If you are building a voice AI, LLM, computer vision, or multimodal system that needs Caribbean or Latin American coverage, this document tells you how HRG collects the data, what quality controls apply, and where our capabilities end.

Procurement and legal teams evaluating vendors. If you are conducting due diligence on HRG as a potential vendor, this document describes our consent framework, chain of custody, data protection compliance, and audit trail structure.

Academic researchers and regional AI programs. If you are conducting AI research on Caribbean or Latin American languages, populations, or use cases, this document describes methodology that may be adaptable to your work under the document's open license.

1.3 What this document is not

This is not a sales pitch. It is a methodology description with honest limits.

This is not an annotation methodology. HRG collects field data. We do not run annotation pipelines. Buyers needing collection plus large-scale annotation should read Section 6 below on honest limits.

This is not a research paper. This is an operating document from a practitioner firm. Claims are supported by references where public sources exist. Operating specifics are drawn from HRG's 40+ years of Caribbean and Latin American field research.

1.4 Version history

Version 1.0 (this document): initial publication, September 14, 2026.

2. Regional context

2.1 Why Caribbean and Latin American field collection is structurally different

Field data collection in the Caribbean and Latin America differs from collection in North American or European markets in four ways.

Language distribution is under-represented in training corpora. Spanish and Portuguese together represent roughly 2 to 3 percent of the training material behind existing AI models, per analysis published by TechPolicy.Press summarising figures from Chile's National Center for Artificial Intelligence (CENIA). Caribbean creoles including Jamaican Patois, Trinidadian English Creole, Papiamento, and Haitian Creole are documented as systematically under-represented across mainstream multilingual NLP benchmarks by the CreoleVal work (Lent et al., 2024, Transactions of the ACL).

Demographic distribution differs from evaluation datasets. Commercial facial recognition and computer vision evaluation datasets are built primarily on North American and European populations. Afro-Caribbean, Indo-Caribbean, and mixed-heritage populations are absent from most evaluation datasets at the discrete-subgroup level, per the NIST Face Recognition Vendor Test evaluations and Buolamwini and Gebru's Gender Shades work.

Data protection frameworks vary by country and change more rapidly. Jamaica's Data Protection Act 2020 was proclaimed into force in 2023. Trinidad and Tobago's Data Protection Act 2011 is under phased implementation. Dominican Republic's Ley 172-13 took effect in 2013. Field collection in these markets requires country-specific compliance rather than a single regional framework.

Field infrastructure requires named country coordinators. Remote crowdsourcing does not scale reliably for country-specific demographic quotas across these markets. Community recruitment, venue access, and quality assurance require on-the-ground presence in each market.

2.2 The training data gap in one paragraph

Sovereign AI programs are addressing the gap at the language-model layer. Latam-GPT, launched February 2026 by CENIA with 15 partner countries and over 200 collaborators, is training the first Latin American language model on 8+ TB of curated Spanish and Portuguese corpora. Brazil's PBIA 2024-2028 allocates R$1.1 billion specifically to a Brazilian Portuguese LLM. What these programs do not solve is field-level data collection infrastructure for downstream ML projects. That is where field vendors like HRG fit.

For a fuller treatment of the gap, see the analysis at hoperesearchgroup.com/blog/caribbean-latam-ai-training-data-gap.

3. The five-stage workflow

Every HRG field collection engagement runs through five stages. This section describes each stage in enough detail to evaluate whether HRG's process fits a specific brief.

3.1 Stage 1: brief scoping and feasibility

Objective. Determine whether a proposed collection is achievable at the required volume, quotas, timeline, and budget before either party commits.

Trigger. A client or aggregator sends a brief describing the collection scope. Typical brief elements include target market or markets, modality (speech, video, image, text), volume, demographic quotas, dialect requirements, timeline, and price point.

Activities. HRG's country office or coordinator team reviews the brief against on-the-ground reality in each target market. Specifically:

Volume feasibility. Can the target volume be sourced in the target markets within the timeline? Where volume exceeds what a single market can produce, we split across markets or extend the timeline.
Demographic feasibility. Can the required demographic quotas be filled in the target markets? Some quotas (Muslim community in Jamaica, Indo-Trinidadian participants in Barbados) are constrained by underlying demographics. Better to surface at scoping than during field work.
Dialect and language feasibility. Some briefs specify dialects at a granularity our recruitment channels cannot verify (specific parish-level Jamaican Patois variants, specific Bhojpuri sub-dialects in Trinidad). We flag these at scoping.
Timeline feasibility. Field work has inherent variability. Weeks with holidays, weather events, or local election periods reduce recruitment capacity. We factor these in at scoping.
Budget feasibility. Some markets have higher field costs. Trinidad runs 20 to 25 percent above Suriname or Dominican Republic on participant incentives, venue costs, and coordinator time. We flag this against the client's rate card at scoping rather than at delivery.

Output. A feasibility assessment returned within 48 to 72 hours of receiving the brief. The assessment states either (a) the brief is achievable as specified, (b) the brief is achievable with specific modifications, or (c) HRG is not the right vendor for this brief.

3.2 Stage 2: recruitment infrastructure

Objective. Build the participant pipeline required to meet the brief's volume and quota targets.

Activities. HRG's country coordinators recruit through four channels.

Community networks. Community-based organisations, sports clubs, civic groups, and neighbourhood associations provide access to specific demographic segments. Community recruitment is our default first channel because incentive alignment is strongest, referral quality is highest, and participant no-show rates are lowest.

University partnerships. University of the West Indies (Mona, St Augustine, Cave Hill campuses), University of Technology Jamaica, Universidad Autónoma de Santo Domingo, Pontificia Universidad Católica Madre y Maestra, INTEC, Universidad de los Andes, and regional university networks provide access to young adult participants, multilingual participants, and specialized language communities.

Religious institutions. For specific religious communities (Muslim, Hindu, Christian denominations), religious institutions are typically the highest-yielding recruitment channel. We work with the specific institution's leadership on informed recruitment aligned with the community's own norms.

Vetted online recruitment. For scale beyond community and institutional networks, we use vetted online recruitment channels with our own screening pass rather than pure crowdsourcing platforms.

Over-recruitment. For remote self-record briefs, HRG over-recruits against the target by 30 to 40 percent to absorb drop-off, no-shows, technical failure, and QA rejections. For on-site collection, over-recruitment sits at 15 to 20 percent because coordinator oversight catches issues earlier.

Output. A verified participant pipeline sized to meet Stage 3 collection volume with quota margin.

3.3 Stage 3: field collection execution

Objective. Capture the actual data records against the brief's specifications.

Activities. Field execution varies by modality.

Speech and voice. Participants record on either the client's proprietary recording app, an HRG-provided recording app, or through a browser-based portal. Recording specifications (typically 48kHz WAV, 16-bit, with transcript and speaker metadata) are documented per brief. Participants are given a brief of the required speech content: read passages, spontaneous responses to prompts, dialogue with an interlocutor, or scripted scenarios. They record within the specified constraints.

Video and computer vision. Participants record video segments per brief (typically 720p or higher, 30fps, .mp4 or .mov). Video briefs specify duration, framing, lighting, and any accessory or activity requirements. HRG coordinators verify setup before submission for on-site collection. For remote self-record, coordinators verify submissions against spec in first-pass QA.

Image. Static image capture per brief specifications. Includes headshot capture for facial data, product-in-hand images for retail computer vision, and environmental images for scene classification.

Text. Written responses per brief. Includes prompt responses, dialogues, translation pairs, evaluation traces, and structured question-and-answer content. Delivered with contributor attribution and dialect flags where the brief requires them.

Consent capture. Every participant signs a project-specific consent form (see Section 4) before recording. Consent capture is either via signed physical form archived to the consent-record ledger, or via digital consent flow if the client's platform captures consent directly at recording time.

Coordinator oversight. For on-site collection, an HRG coordinator is present to verify participant identity against pre-registration, verify consent capture, verify recording setup meets spec, and conduct first-pass QA before submission.

Output. Submitted records that meet the brief's specifications and have documented consent.

3.4 Stage 4: quality assurance and rejection handling

Objective. Filter out records that do not meet the brief's specifications before they reach the client.

Activities. First-pass QA runs against brief specifications. Typical QA checks include:

Specification compliance. Recording length, sample rate, bit depth, frame rate, resolution, aspect ratio, format.
Content compliance. Correct read passage or prompt response, complete answer, appropriate language or dialect, no synthetic or AI-generated content.
Participant compliance. Verified participant identity, no duplicate participation, consent captured before recording began.
Environmental compliance. Background noise, lighting, framing, absence of unauthorized third parties in frame.

Records that fail QA cycle back to the participant for re-record where the failure is correctable (audio too quiet, participant misread the prompt, wrong framing). Records that fail for uncorrectable reasons (duplicate participant, ineligibility discovered post-recording, consent gap) are replaced from the over-recruited pool built in Stage 2.

Rejection categorisation. When rejecting a record, we tag the rejection reason from a standard category list: specification failure, metadata mismatch, incomplete consent, duplicate participant, reused background, collection against closed quota. This lets us train the field team from rejection patterns week to week, and gives the client a documented rejection log they can reconcile against their own downstream QA.

Output. A cleared record set that meets brief specifications, with rejection log for the batch.

3.5 Stage 5: delivery and handoff

Objective. Deliver the collected records to the client's ingestion pipeline with documented provenance.

Activities. Delivery format varies by client. Common patterns:

Direct upload to client's cloud storage (S3, GCS, Azure Blob) with pre-shared credentials.
Upload to client-provided ingestion portal (typical for aggregator engagements).
Batch delivery via secure file transfer with cryptographic hashes for verification.

Every delivered record includes standard metadata:

Participant identifier (anonymized in delivered data, linked to consent record via HRG's consent-record ledger).
Demographic metadata per brief specification.
Dialect and language flags.
Recording specifications (device type where relevant, timestamp, location if collected per brief).
Consent form reference.

Consent-record ledger update. HRG maintains a separate consent-record ledger cross-referencing each participant identifier to their signed consent form. On delivery, ledger entries for the batch are marked as delivered and locked against modification.

Data deletion schedule. Following the contract-specified retention period, HRG deletes source records from working systems while preserving the consent-record ledger for its full statutory retention period per applicable data protection law.

Output. Delivered records in the client's ingestion pipeline, closed batch delivery report, closed consent-record ledger entries.

4. Consent chain and data protection

4.1 Consent framework

Every participant in an HRG field collection signs a project-specific consent form that names:

The collector. Hope Research Group (or the specific HRG entity depending on jurisdiction).
The data controller. The client commissioning the collection.
The intended use of the data. Specific to the project: AI model training, evaluation, benchmark contribution, research.
The retention period. Contract-specified.
The participant's right of withdrawal. How and where to exercise it.
Cross-border data transfer notice. If applicable to the target jurisdictions.

Consent forms are language-appropriate for the market (English, Spanish, Portuguese, Dutch, or French depending on jurisdiction).

Consent must be captured before recording begins. Records without documented consent are not delivered.

4.2 Consent-record ledger

The consent-record ledger is HRG's central archive of participant consent. It contains:

Participant identifier (unique per collection engagement, anonymized in delivered data).
Consent form reference (physical or digital).
Timestamp of consent capture.
Withdrawal record if the participant exercises their right of withdrawal.
Retention period end date.

The ledger is queryable per record. An audit trail can be produced for any single participant on request from the client, or from a regulator with jurisdiction.

The ledger is stored separately from the delivered records themselves, with restricted access to designated HRG personnel only.

4.3 Data protection framework by market

Data protection compliance varies by country of collection. HRG operates in compliance with the applicable framework in each market:

Jamaica: Data Protection Act 2020, proclaimed into force December 2023.
Trinidad and Tobago: Data Protection Act 2011, under phased implementation through proclamation of specific sections.
Dominican Republic: Ley 172-13 (2013).
Brazil: Lei Geral de Proteção de Dados (LGPD, 2018).
Argentina: Ley 25.326 Personal Data Protection Act (2000, subsequently amended).
Mexico: Ley Federal de Protección de Datos Personales en Posesión de los Particulares (2010).
Colombia: Ley 1581 de 2012 (Personal Data Protection Law).
Chile: Ley 19.628 sobre Protección de la Vida Privada (1999, subsequently amended).

Other Caribbean and Latin American markets fall under either specific national frameworks or, where absent, HRG applies international best practice (GDPR-equivalent standards) as a working default.

4.4 Cross-border transfer

Many collection engagements involve cross-border transfer of collected data from HRG's country of collection to the client's country of processing. HRG's consent forms include explicit cross-border transfer notice where applicable, and HRG operates within the applicable data-protection framework's provisions for such transfers.

4.5 Chain of custody

For every delivered record, HRG can produce a chain of custody documenting:

The recording device or platform used.
The recording timestamp.
The participant identifier.
The consent form reference.
The QA pass timestamp.
The delivery timestamp.

Chain of custody is a matter of internal recordkeeping rather than blockchain or notarized attestation. Specific chain-of-custody documentation can be provided under contractual audit rights.

5. Demographic and dialect flagging

5.1 Demographic metadata

Every delivered record includes demographic metadata per the brief's specification. Standard fields include:

Age band (typically 18-24, 25-34, 35-44, 45-54, 55+).
Gender (self-reported: male, female, other, prefer not to say).
Skin tone (self-reported or coordinator-assessed against the Monk Skin Tone scale, where relevant to brief).
Language or languages spoken (primary language, additional languages).
Region of residence (country plus sub-national region).

Additional fields per client specification (education level, occupation category, religious identification, ethnic identification) are captured only when the brief specifies them and the participant consents to their capture.

5.2 Dialect flagging

Dialect flags are critical for Caribbean creole and regional Spanish briefs because bundled "LatAm Spanish" or "Caribbean English" categories fail to reflect real linguistic diversity that affects downstream model performance.

Standard dialect flags applied per record where relevant:

Jamaican briefs:

Jamaican Patois (Jamaican Creole English).
Jamaican Standard English.
Code-switching between the two.

Trinidadian briefs:

Trinidadian Standard English.
Trinidadian English Creole.
Trinidadian Hindi or Bhojpuri variant (for Indo-Trinidadian participants where relevant to brief).

Dominican Spanish briefs:

Dominican Spanish (with feature flags for aspiration or elision of syllable-final /s/, yeísmo, and other salient features where specified).
Standard Spanish variant.

Multi-country LATAM Spanish briefs:

Mexican Spanish.
Rioplatense Spanish (Argentina, Uruguay).
Chilean Spanish.
Andean Spanish (Peru, Bolivia, Ecuador highlands).
Caribbean Spanish (Dominican Republic, Cuba, Puerto Rico).
Colombian Spanish (with sub-regional flags where specified).

5.3 Handling code-switching

Caribbean speakers frequently code-switch between creole and standard forms within a single utterance or across dialogue turns. Handling depends on brief specification:

If the brief wants pure register samples (Patois only, Standard English only), the recording script or prompt is designed to elicit the target register, and participants are asked to hold the register throughout.
If the brief wants natural speech including code-switching, participants respond naturally and each utterance or turn is flagged per its dominant register.

6. Honest limits

6.1 What HRG does not do

HRG does not run large-scale annotation pipelines. If a project needs field collection plus downstream annotation at scale (bounding boxes, transcription, semantic tagging), HRG collects and hands off to the client's annotation partner. We can recommend annotation partners on request but do not sub-contract annotation.

HRG does not train AI models. We collect the data that models are trained on.

HRG does not sell pre-collected datasets. Every collection is commissioned fresh for a specific client engagement.

HRG does not host client models or platforms. Delivery is direct to the client's ingestion pipeline.

6.2 Where HRG is not the right vendor

End-to-end training data pipelines with annotation at scale. Better vendors: TELUS Digital, Sama, iMerit, Scale AI, Appen, where end-to-end integration with in-house annotation matters more than country-vendor field presence.

Single-market small-batch collection outside HRG's 17 markets. HRG covers the Caribbean and Latin America. For African-market speech, Southeast Asian dialect data, or Indian regional language collection, we are not the right vendor.

Real-time or 24/7 conversational data collection. HRG's collection is batch-oriented over days-to-weeks project timelines. Real-time conversational data collection requires infrastructure we do not operate.

Speaker verification or biometric authentication training data at scale. Feasible on our infrastructure but not our specialty. Better to work with vendors specialized in biometric collection.

6.3 Where quotas can be constrained

Some briefs specify demographic quotas that are constrained by underlying demographics in HRG's markets. Examples:

Muslim community briefs in Jamaica. Jamaica's Muslim population is approximately 1 percent of total. Brief-scale Muslim quotas are not achievable in Jamaica alone. Better markets: Trinidad and Tobago (approximately 5 to 6 percent Muslim per the 2011 census), Guyana, Suriname.

Indo-Caribbean quotas outside Trinidad and Guyana. Indo-Caribbean populations are concentrated in Trinidad (approximately 35 percent per 2011 census) and Guyana. Elsewhere in the Caribbean, Indo-Caribbean quotas are constrained by underlying demographics.

Non-Christian religious identification briefs. Caribbean and Latin American populations are predominantly Christian. Non-Christian religious identification briefs at volume need specific market selection.

We flag these constraints at Stage 1 feasibility rather than at Stage 4 QA.

7. Version and license

Version: 1.0

Date: September 14, 2026

Author: Kurt Wedderburn, Managing Director, Hope Research Group

License: Creative Commons Attribution 4.0 International (CC-BY 4.0)

This document is published under CC-BY 4.0. You are free to share, adapt, and build upon this work for any purpose, including commercial use, provided you give appropriate credit to Hope Research Group.

References

CENIA / Latam-GPT publications: latamgpt.org
TechPolicy.Press analysis of Latam-GPT and regional AI corpus statistics
Lent, H. et al. 2024, "CreoleVal: Multilingual Multitask Benchmarks for Creoles," Transactions of the Association for Computational Linguistics, vol. 12, pp. 950 to 978. DOI 10.1162/tacl_a_00682
Buolamwini, J. and Gebru, T. 2018, "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification," Proceedings of Machine Learning Research 81
NIST Face Recognition Vendor Test (ongoing evaluations)
Governo do Brasil (MCTI), 2024, "Plano Brasileiro de Inteligência Artificial (PBIA) 2024-2028"
CENIA, February 2026, Chile-Panama bilateral AI cooperation announcement
Hope Research Group, 2026, "The Caribbean and Latin American AI Training Data Gap," hoperesearchgroup.com/blog/caribbean-latam-ai-training-data-gap

Contact

Kurt Wedderburn, Managing Director admin@hoperesearchgroup.com +1 954 862 3661

Hope Research Group hoperesearchgroup.com | Book a consultation

Downloads last month: 17

Total file size:

28.3 kB