Human data · Latin America

Latin American human data, with a person behind every record.

Pluri runs its own collection infrastructure across the region: speech, imagery and real-world task data, captured by identity-verified contributors, validated by automated QA and human review, and paid on local rails from Pix to SPEI. Every delivery ships with per-record provenance.

record ACCEPTED

    the provenance pipeline, live · illustrative values

    1.4M+identified people in the owned panel
    20+cities with active field operations
    2languages collected natively · pt-BR and es-419
    2019running field data operations since

    The gap

    Latin America is the most underweighted major region in training corpora.

    700M+native speakers of Spanish and Portuguese worldwide
    ~6%combined share of both languages in global web text corpora

    Published research on Portuguese language models shows that the gain from native-language pretraining comes from cultural and domain knowledge, not grammar. The same logic holds across the region. What models are missing is not syntax. It is Latin American context: how people speak, pay, shop, complain and fill out forms.

    That context does not exist in scraped web data. It has to be collected from people: with regional accent balance, real devices, real environments and documented consent.

    Pluri exists to supply it, at contract grade.

    Capabilities

    Three modalities, one collection machine.

    Custom collection against your spec. Fresh, task-generated data from real contributors. Nothing scraped, nothing synthetic.

    MOD·01

    Speech collection

    Scripted prompts, spontaneous conversation, wake words, code-switching, noisy environments. Accent quotas balanced across the region, in Brazilian Portuguese and Latin American Spanish.

    • 48kHz WAV or client spec
    • demographic attributes per speaker
    • pt-BR and es-419 · accent quotas enforced at recruiting
    • optional verbatim transcription
    MOD·02

    Imagery and documents

    Real-world photos, receipts, screens, shelves and street scenes. Captured on the contributor's own device, validated at submission time.

    • PII redaction at ingestion, logged per record
    • device and capture metadata preserved
    • perceptual-hash dedup across full history
    • human review gate on every batch
    MOD·03

    Task and app-flow data

    Guided real-world tasks on Latin America's digital rails: instant payments, delivery, e-commerce, government services. Step-level capture, with consent.

    • step-level screen capture, redacted
    • instruction-compliance checks per step
    • flows that exist in no scraped corpus
    • targeting by country, city and profile

    Industries

    Ground truth by vertical, from the consumer's own screen.

    Datasets and standing panels are scoped per industry. What we capture is what the market actually shows people, and what people actually do about it.

    FIN

    Financial services

    Fees and rates as displayed, credit offers received, onboarding and payment flows across banks and fintechs.

    • fees and offers as shown to real users
    • instant-payment and checkout flows
    • onboarding funnels, step by step
    DLV

    Delivery and foodtech

    Order receipts, baskets and prices paid, promos seen, delivery times as experienced across delivery apps.

    • order-level receipts with prices paid
    • promo and coupon exposure
    • delivery time as experienced, not as promised
    ECOM

    E-commerce and marketplaces

    Search results, product pages, carts and checkouts as shown to real shoppers, across competing marketplaces.

    • search ranking as displayed
    • price and coupon at checkout
    • cross-marketplace comparison, same shopper
    RET

    Retail and CPG

    Shelf and street photography, in-store price checks and paper receipts from physical retail, digitized and validated.

    • shelf and planogram photos
    • in-store price capture
    • receipts digitized line by line
    MOB

    Mobility

    Fare quotes, surge behavior and wait times across ride-hailing and mobility apps, captured on both sides of the trip.

    • fare quotes as displayed
    • surge and wait-time evidence
    • rider-side and driver-side capture
    MED

    Media and subscriptions

    Plan and price screens, paywalls, churn and win-back offers across streaming and subscription products.

    • plans and prices as displayed
    • paywall and offer variants
    • cancellation flows, step by step

    Provenance

    Every claim about quality reduces to one question: show me a record.

    This is what one delivered unit looks like. Not a summary metric over the dataset: the audit trail of a single record, the way it ships.

    record plr_a91c04…7e4 SPEECH · PT-BR ACCEPTED

    Media

    type
    audio · wav · 48kHz · mono
    duration
    00:41
    task
    spontaneous_speech · prompt v3.2
    environment
    indoor, ambient noise present (as specified)

    Contributor

    id
    c_7f20…9b (pseudonymous)
    profile
    woman · 25 to 34 · Ceará · Northeast accent
    identity
    name, phone and payment identity on file, checked at payout
    history
    61 accepted units · 0 fraud flags

    Consent

    version
    v2.1 · accepted 2026-07-14 09:12 BRT
    scope
    commercial model training · worldwide · sublicensable
    withdrawal
    honored forward, logged

    Quality

    auto QA
    6/6 checks passed (authenticity, dedup, spec match)
    human review
    approved · reviewer r_04 · 2026-07-15
    redaction
    1 region masked · log retained

    Payment

    payout
    Pix · confirmed 2026-07-18
    condition
    released on acceptance, never before

    Counterparty

    operator
    registered company in the region, tax ID on contract
    support
    contributor support in Portuguese and Spanish, with SLA

    Ask any vendor for this view of one record. That question is the entire audit.

    Field values above are illustrative. The structure is contractual.

    Pipeline

    One machine: reach, execute, verify, pay.

    The same four-stage pipeline runs every modality. Change the task, keep the controls.

    01 / REACH

    Targeted recruiting

    Recruiting from an owned, identified contributor base, filtered by country, city, demographic and accent quota. No marketplaces, no scraping, no anonymous crowd.

    02 / EXECUTE

    Guided capture

    Contributors complete the task in a conversational collector that validates at submission time: wrong city, wrong format or missing fields never enter the pipeline.

    03 / VERIFY

    Machine plus human QA

    Automated checks for authenticity, spec match and cross-contributor duplication, then a human review gate on every batch. Uncertain units go to quarantine, not to you.

    04 / PAY

    Acceptance-linked payout

    Contributors are paid on local instant rails, Pix and country equivalents, after client acceptance, with tax reporting handled. Paid contributors return. Returning contributors are the quality flywheel.

    Continuous panels

    Standing communities that turn screens into a time series.

    The same machine, left running. Consented cohorts of 500 to 2,000 people per industry submit screen-level evidence from their own devices every month: prices paid, offers seen, flows completed. Longitudinal, verified, redacted.

    What subscribers get

    A monthly delivery combining the record-level dataset and an aggregated read: price and promo movements, share shifts, flow changes. Same cohort over time, so a movement is a movement, not sampling noise.

    • monthly drop · dataset plus aggregated read
    • every record ships with full provenance
    • same cohort tracked across waves
    • PII redacted before anything leaves the pipeline

    How it is sold

    Syndicated by default: every subscriber receives the same deliverable, and no one can commission collection aimed at a specific company. A separate license makes the underlying screens available for model training.

    • syndicated subscription, no exclusives
    • no client-targeted collection, ever
    • optional AI-training license on the corpus
    • panel consent covers both uses explicitly

    Proven in production: our standing screen-collection program runs at a contracted volume of 20,000 validated units per month.

    For data platforms

    Built to slot into your delivery pipeline.

    If you sell data to AI labs and need Latin American supply that survives your QA bar, this is the product.

    Project delivery

    Your spec, our field. Send collection requirements and get a costed pilot proposal within 72 hours. Daily batch manifests, verdicts returned on your template, rejected units replaced.

    • costed pilot proposal in 72h
    • daily batch close with manifest
    • client verdict template, reconciled per unit
    • white-label delivery under NDA

    Reserved capacity

    Contract a monthly throughput floor by modality, country, city and demographic, with burst on demand. Your roadmap gets guaranteed Latin American supply instead of spot-market risk.

    • monthly throughput floor, by contract
    • priority quotas by country, city and profile
    • burst capacity on demand
    • quarterly quality review with acceptance data

    Our field operation has delivered to global data platforms and research companies, under NDA, since 2019.

    Compliance

    Compliance is the product, not the disclaimer.

    Privacy-law native

    Consent logged with version, timestamp and scope. Purpose-bound collection, opt-out honored and recorded, retention and deletion per contract. Brazil's LGPD as the baseline, aligned with privacy regimes across the region. Data processing agreement as standard.

    Redaction at capture

    Personal identifiers are masked at ingestion, before storage and before delivery, with a per-record redaction log. Uncertain redactions are quarantined for human review, never shipped.

    A real counterparty

    A registered company in the region, enforceable contracts, and contributor support in Portuguese and Spanish. When something needs resolving, there is someone to call. In this market, that is rarer than it should be.

    Fair, documented payment

    Contributors are identified and paid per accepted unit on local rails, with tax reporting handled. Payment terms are stated before the task starts. Disputes have a channel and an SLA.

    Start

    Send a spec. Get a costed pilot in 72 hours.

    One page is enough: modality, volume, demographic targets and your QA bar. We reply with scope, timeline and unit pricing.

    hello@pluri.lat typical first pilot: 2 to 4 weeks, one city, full provenance