---
title: "Data Labs · Marxen"
url: "https://marxen.in/data-labs"
description: "Close to 500,000 verified contributors across India, Nepal and Sri Lanka. Speech, text, image, egocentric and multimodal sensor data, human-collected and consent-documented."
---

# Compute is purchasable. Reach is not.

Close to 500,000 verified contributors across India, Nepal and Sri Lanka. Speech, text, document, image, egocentric and multimodal sensor data, gathered by people who are physically there.

## 01 · Why this is the hard part

Everyone building a model needs the same thing: more data, from more of the world, collected and consented properly. The bottleneck is not annotation software. It is reach.

A clean hour of code-switched Tamil and English speech from a 52-year-old bus conductor needs someone standing in front of that bus conductor, explaining what the recording is for, taking real consent, capturing usable audio in a noisy depot, and paying him the same day.

Our network was built over years for ground activations and campus programmes. It was engineered to put trained people in front of strangers at scale, and we repointed it at data. Contributors are identity-verified, briefed, trained and paid, which is what separates a corpus you can attest to from one assembled by anonymous accounts optimising for volume.

## 02 · What we produce

**Speech and ASR corpora.**
Read, spontaneous and conversational speech. Code-switched Indian-language data, which is how India actually speaks and which most corpora quietly exclude. Field-condition audio with the traffic, machinery and crowds a deployed model will really meet.

**Transcription at volume.**
Multilingual, with speaker separation, timestamps and structured output.

**Text and NLP corpora.**
Indic datasets for fine-tuning across education, healthcare, legal and government work. Annotated by native speakers, never machine-translated labels dressed up as ground truth.

**Image, video and document annotation.**
Bounding boxes, segmentation, classification, OCR ground truth, and document-structure labelling for layouts Western training sets have never encountered, including handwritten registers and multi-script forms.

**Egocentric data.**
First-person capture of people doing real tasks. The scarcest category in the market, and the one physical AI depends on most.

**Multimodal sensor data.**
Time-aligned audio, video, IMU and depth captured together, for models that have to work in a body rather than a browser tab.

**Retrieval-ready knowledge bases.**
Ingestion, chunking and vector-store construction over an institution's own documents.

## 04 · How we work

Human-led annotation with structured QA. Not a crowd marketplace, and not a model labelling its own training data. Every workflow leaves an audit trail: who labelled what, against which version of the guidelines, at what inter-annotator agreement.

Consent is explicit, recorded and revocable, with the terms written in the contributor's own language. Contributors are paid on schedule and told what the data is for. For a regulated buyer that is the difference between a corpus you can train on and one your counsel makes you delete.

## 05 · Who buys this

Frontier and foundation-model labs adding multilingual capability. Indian AI companies fine-tuning on domain data. Enterprises building on their own historical material. Government language-infrastructure programmes. Robotics teams that need egocentric and sensor data from environments that look nothing like a California office.

## Pages

- [Coverage](https://marxen.in/data-labs/coverage.md)