We've been training AI models for global tech giants since 2018 — ride-hailing platforms, social networks, e-commerce, foundation-model labs, and short-form video. Marano operates as a closed, NDA-bound studio: every annotator is recruited, screened, and audited under our own QA spine.
Our linguists, voice talents, and domain reviewers were already the operational backbone of high-stakes localization work. When global tech labs needed RLHF data, voice corpora, and content-moderation taxonomies, we adapted the same workforce — under the same QA spine — to the AI training brief.
That continuity matters. We don't crowdsource. We don't onboard anonymous workers behind a platform UI. Every name on a payroll we issue an NDA for has been calibrated, qualified, and audited — and most have been with us for years.
Reinforcement Learning from Human Feedback (RLHF), red-teaming, prompt-response pair authoring, instruction-tuning datasets, and human-preference comparison ranking for foundation-model labs.
Image, video, document, and 3D point-cloud annotation for computer vision, autonomy, content moderation, and document AI workloads — backed by a 4-stage QA pipeline.
Multi-speaker, multi-domain speech collection inside our own ISO-grade booths. From neutral TTS corpora to spontaneous conversation, code-switching, and paralinguistic emotion sets.
Machine-translation post-editing benchmarks, LLM output evaluation, multilingual NLU/NLG datasets, and culture-aware content review across 40+ language pairs.
Five modality families, each delivered at production volume. Per program we publish throughput targets, IAA bands, and reviewer composition before kick-off so the customer can model unit economics before signing.
RLHF pairs, SFT data, prompt-response sets, MTPE evaluation, NLU intents, toxicity / policy classification
40+ pairs, SEA & EA depth
Up to 100K samples / week
2D / 3D bounding boxes, polygon & semantic segmentation, keypoints, object tracking, content-moderation taxonomies
Locale-aware moderation in 12+ markets
Up to 500K frames / month
TTS corpora, spontaneous dialogue, code-switching, emotion / paralinguistic tagging, wake-word & command sets
TH, ID, VI, MS, TL, JA, KO, ZH, EN + dialects
Up to 1,000 speakers / quarter
LiDAR / radar / camera fusion ground-truth, ADAS / AV scene labels, point-cloud segmentation, driver-monitoring
Spec language: EN / ZH; cabin-voice in 9 langs
Per program; pilots from 10K frames
OCR, layout parsing, table extraction, redaction taxonomies, multi-language document classification
Latin, CJK, Thai, Arabic scripts
Up to 250K pages / month
Every Marano AI program follows the same operational rhythm. Customers can plug in at any stage — brief design, calibration, production, or QA-only — without forcing a workflow rewrite.
Customer labelling guidelines, taxonomies, and edge-case decks are ingested by senior leads. We rewrite ambiguous instructions into testable rules before annotators touch a single sample.
Domain-screened annotators take blind qualification rounds against a gold-standard set. Pass rates are reported per cohort; only calibrated reviewers enter production.
Work is dispatched in batches sized to the program's throughput target. We hold daily stand-ups with team leads and weekly calibration syncs with the customer.
Dual annotation, senior adjudication, and statistical sampling run on every batch. Inter-annotator agreement (IAA / Cohen's κ) is published per delivery.
Edge-case rulings are looped back into the live guideline. Annotators are re-calibrated whenever the spec moves, so historical batches stay consistent with current intent.
Customer-facing dashboards expose batch status, IAA, and reviewer identity (pseudonymous) so the customer can audit any sample back to its annotator and adjudicator.
Sampling rates, IAA reporting, and audit cadence are configurable per program. We publish per-batch agreement metrics so customers can monitor data drift in real time.
Domain-screened annotators sign NDAs, complete project-specific calibration, and pass blind qualification rounds before touching production data.
Every sensitive sample is independently annotated by two specialists. Disagreements are routed to senior reviewers, never silently averaged.
Domain leads adjudicate edge cases against the customer's labelling guideline and feed clarifications back into the live spec.
Sampled audits with inter-annotator agreement (IAA / Cohen's κ) reporting per batch, plus customer-facing dashboards for traceability.
We share voice infrastructure with our broadcast dubbing operation. The same booths that record anime and theatrical titles capture multi-speaker AI corpora — at 24 / 48 kHz, with demographic and dialectal balancing across speaker pools.
Most of our AI work runs under partner-LSP NDAs. We can't name the customer here, but the verticals, scope, and operational outcomes are real. Detailed program references available under a mutual NDA.
RLHF & safety alignment data for an in-app conversational assistant covering Southeast Asian markets.
2M+ preference pairs, TH / ID / VI / EN, 18-month rolling engagement under partner LSP NDA.
IAA ≥0.82 sustained across batches; spec drift caught and rolled back twice without customer escalation.
Multilingual LLM evaluation — factuality, fluency, and culture-aware critique across SEA + EA pairs.
30K+ structured critiques per quarter, 9 target languages, senior reviewers from media, legal, and medical pools.
Customer's published model cards cite the SEA evaluation set our linguists contributed to.
Content-moderation taxonomy refinement and policy-aware video segmentation for a global UGC product.
500K+ video clips reviewed, 12 locales, dual annotation + senior adjudication on every sensitive sample.
Reviewer attrition <8% / year; policy changes propagated to live spec inside one calibration cycle.
We do not require customers to migrate to a Marano-owned platform. We staff inside the customer's preferred tooling — managed platforms, open-source labelling tools, or custom pipelines — and run our QA spine on top.
We staff and manage workforces inside Scale AI, Appen, Toloka, Lionbridge, RWS / TrainAI and other partner-managed environments.
CVAT, Label Studio, Doccano, Prodigy, V7, Encord — configured per project, with reviewer roles mirrored to our internal hierarchy.
Direct ingest from customer S3 / GCS / Azure Blob, signed-URL workflows, and webhook delivery for batched outputs.
Closed network annotation rooms, mandatory NDAs, device-level controls, and per-program access logs. We can operate inside the customer's VDI when required.
The hardest part of frontier AI work isn't the labelling — it's making sure the same person who labelled batch zero is still making the same call on batch three hundred. That's what our QA spine is for.
Selected engagements delivered directly or through Leading Language Service Providers under strict NDA. Detailed program scopes available on request.
Multilingual, multi-domain, production-volume work where the quality bar is non-negotiable. Pilots can start within two weeks.