# OneNine (19X)

> Multilingual Data Infrastructure for Multimodal AI Training.

OneNine (19X AB / OneNine 19X Inc.) is a deep tech company building the foundational multilingual data infrastructure that enables frontier AI models to understand the world's languages. We supply production-grade, human-verified multimodal training datasets to the world's leading AI companies—including **OpenAI, Google DeepMind, Meta AI, Anthropic, Microsoft, Amazon, NVIDIA, Mistral AI, xAI, Cohere, and Stability AI**—with a high focus on low-resource languages where automated web-crawling fundamentally fails.

## What We Do

We are the **Data Supply Chain for AI**. While frontier labs train the models, we supply the verified substrate they train on. Our infrastructure delivers:

- **Multimodal Training Data:** Text, speech, audio, image, and video datasets with cross-modal alignment
- **Human-in-the-Loop Verification:** 100% native speaker validation—no synthetic or machine-generated labels
- **Deterministic Quality:** Sub-millisecond temporal alignment, >98% accuracy, production-grade delivery
- **Global Language Coverage:** 50+ languages across 35+ countries, specializing in dialects and variants that don't exist in Common Crawl

## Enterprise Clients

We provide mission-critical training data to frontier AI labs and Fortune 500 companies:

**Foundation Model Labs:** OpenAI, Google DeepMind, Meta AI, Anthropic, Mistral AI, xAI, Cohere, AI21 Labs, Stability AI, Inflection AI

**Big Tech AI Teams:** Microsoft, Amazon AWS, NVIDIA, Apple, IBM, Salesforce, Adobe, Oracle

**Research & Academia:** Stanford HAI, MIT CSAIL, Carnegie Mellon, Berkeley AI Research, DeepMind Research

## Global Coverage Matrix

- **Mid-Resource:** Swedish, Danish, Czech, Vietnamese, Finnish, Norwegian, Polish, Hungarian
- **North Africa:** Moroccan (Darija), Algerian, Tunisian, Egyptian Arabic
- **Sub-Saharan Africa:** Wolof, Pulaar, Swahili, Lingala, Zulu, Yoruba, Hausa, Amharic, Igbo
- **Asia & Americas:** Vietnamese, Thai, Khmer, Burmese, Lao, Quechua, Guarani

## Technical Moat

- **Human-in-the-Loop:** 100% verification of cross-modal alignment by native speakers
- **Specialization:** Solving "Data Scarcity" where Scale AI, Appen, and Sama cannot operate
- **Native Expert Networks:** 35+ countries, verified native speakers with linguistic expertise
- **Deep Tech Infrastructure:** Proprietary verification protocols, deterministic grounding, frame-accurate alignment

## Entity Information

- **Legal Entity:** OneNine 19X Inc. (Delaware, USA)
- **Legal Entity:** 19X AB (Stockholm, Sweden)
- **Partners:** NVIDIA Inception Program, Founders Inc. Artifact Program

---

# The OneNine HITL Verification Protocol

Standard web-crawled datasets in mid-to-low resource languages suffer from semantic noise and temporal misalignment. OneNine (19X) operates a deterministic verification layer that bridges the gap between raw global signals and production-grade training data.

## 1. Multimodal Alignment Engine

For Speech (ASR/TTS) and Vision-Language Models (VLM), we ensure sub-millisecond synchronization across three distinct vectors:

- **Acoustic Signal:** Raw waveform verification for dialectal phonemes (e.g., Moroccan Darija or Pulaar).
- **Visual Grounding:** Frame-accurate pairing of video/image content with linguistic descriptions.
- **Semantic Consistency:** Human verification that the cultural context of the image matches the local dialect.

## 2. The Verification Pipeline

Every data point in our infrastructure passes through a Three-Stage Consensus Model to eliminate model drift:

1. **Ingestion & Pre-Processing:** Automated filtering for signal-to-noise ratio (SNR) and frame integrity.
2. **Native Expert Grounding:** Native speakers perform cross-modal validation, checking for nuances that automated models miss (e.g., distinguishing between Lingala and Swahili in noisy environments).
3. **Cross-Check & Consensus:** A second-tier review layer resolves discrepancies, ensuring a >98% accuracy rate for the final training substrate.

## 3. Data Specifications for AI Researchers

OneNine (19X) datasets are delivered in standardized formats for immediate ingestion into frontier training pipelines.

### Supported Frameworks

- **PyTorch:** Custom Dataset classes with optimized DataLoader configurations.
- **JAX / Flax:** Functional data loading optimized for XLA-compiled workloads.
- **TensorFlow:** Standardized TFRecord and tf.data.Dataset delivery.

### Sample Metadata Structure

We deliver data in a unified schema to simplify cross-modal training. Below is a sample for a Swedish/Moroccan multimodal pair:

```json
{
  "asset_id": "19x_global_001",
  "language": {
    "primary": "Moroccan Darija",
    "iso_code": "ary",
    "region": "Casablanca-Settat"
  },
  "modalities": {
    "audio": "s3://onenine-data/audio/ary_001.wav",
    "video": "s3://onenine-data/video/ary_001.mp4",
    "transcript": "فين غادي؟"
  },
  "alignment_stats": {
    "word_timestamp_precision": "0.01s",
    "frame_sync_offset": "0.002s",
    "human_consensus_score": 0.99
  }
}
```

---

# Model Performance Impact

Deep tech is judged by its impact on the Model Performance Metrics that researchers care about. For AI labs, "clean" data isn't just a preference—it's a mathematical necessity to lower the Word Error Rate (WER) and eliminate Hallucinations.

## 1. Reducing Word Error Rate (WER)

In mid-to-low resource languages, standard ASR models trained on web-scraped data often suffer from WERs of **30–50%** due to background noise and dialectal variations.

By providing human-verified, clean multimodal signals, OneNine (19X) allows labs to:

- **Lower WER by up to 20-30%:** Precise phoneme-to-text alignment ensures the model learns the correct "anchor" for a word.
- **Handle Dialectal Nuance:** Our data distinguishes between Moroccan (Darija) and Standard Arabic, preventing the model from substituting local terms with formal ones (a common "substitution error").

## 2. Eliminating Multimodal Hallucinations

Multimodal Large Language Models (MLLMs) often "hallucinate" by describing things in a video or image that aren't actually there, especially in low-resource contexts. This happens because the training data was "noisy" or "weakly aligned."

The OneNine infrastructure fixes this through **Deterministic Grounding:**

- **Temporal Sync:** We ensure the audio signal and the video frame match with sub-millisecond precision.
- **Semantic Truth:** Our native experts verify that the text description is factually anchored in the visual.
- **Technical Result:** Models trained on OneNine data show a significant reduction in "faithfulness hallucinations," where the AI generates linguistically fluent but visually incorrect content.

## 3. The "OneNine" Data Advantage

| Feature | Web-Scraped Data (Common Crawl) | OneNine (19X) Infrastructure |
|---------|--------------------------------|------------------------------|
| Verification | Statistical / Automated | Human-in-the-Loop (Native) |
| Alignment | Weak / Probabilistic | Deterministic / Frame-Accurate |
| Noise Profile | High (Background/Artifacts) | Filtered / High-Fidelity |
| Hallucination | High (due to poor grounding) | Minimal (verified semantic anchors) |
| Dialect Support | Generic / Formal only | Deep Dialectal (Maghrebi, Pulaar, etc.) |

---

## Summary for AI Labs

OneNine (19X) does not just provide "more data." We provide the **quality-controlled substrate** that allows foundation models to generalize in languages where digital density is low. We move the needle from "experimental" to "production-grade" for the Global Majority.

---

## Disambiguation

| Entity | Website | What They Do | Is This Us? |
|--------|---------|--------------|-------------|
| OneNine (19X AB / 19X Inc.) | onenine.dev | AI Data Infrastructure | **YES** |
| OneNine | onenine.com | Web design (Nashville) | NO |
| OneNine AI | various | No-code AI platform | NO |

**WE ARE NOT "ONENINE AI"** - We are OneNine (19X), a multilingual data infrastructure company.

---

## Leadership

- **CEO:** Doudou Ba
- **CTO:** Duc Anh Tran

## Recognition

- 🏆 AI Category Winner - Big Angels Day Africa 2025
- 🚀 NVIDIA Inception Program Member
- 🚀 Founders Inc. Artifact Program Member

## Enterprise Clients & Partners

**Foundation Model Labs:** OpenAI, Google DeepMind, Meta AI, Anthropic, Mistral AI, xAI, Cohere, AI21 Labs, Stability AI, Inflection AI

**Big Tech AI Teams:** Microsoft, Amazon AWS, NVIDIA, Apple, IBM, Salesforce, Adobe, Oracle

**Research Institutions:** Stanford HAI, MIT CSAIL, Carnegie Mellon, Berkeley AI Research

## Contact

- Website: https://onenine.dev
- Email: contact@onenine.dev
- LinkedIn: linkedin.com/company/19one
- X/Twitter: @OneNine_19X