Auditable data labeling services
Data labeling and annotation services for text, documents, images, audio and structured data, checked against written guidelines. Every batch ships with a quality report you can hand to an auditor.
What data labeling involves
Data labeling turns raw text, images, audio or records into examples a model can learn from or be tested against. The hard part isn’t the clicking. It’s writing guidelines tight enough that two labelers give the same answer, and then checking that they do.
That’s why every project with us starts with a pilot batch. We measure how often labelers agree, fix the guidelines where they don’t, and only then scale up. You see the numbers at each step.
Human-in-the-loop data for models you have to stand behind
Annotation for text, documents, images, audio and structured data, done by trained labelers and checked in layers. Each batch comes with a quality report, not just a file.
Built for production models
Our processes are set up for real machine learning, LLM evaluation and RAG systems, where a bad label shows up later as a wrong answer in front of a customer.
Teams that scale with the backlog
From a small pilot to an ongoing labeling team with QA reviewers and a coordinator.
Quality on the record
Written guidelines, layered review and regular reports, so you can show how a dataset was made when someone asks.
What you can use our data labeling for
Classic machine learning, and the newer work around LLMs and retrieval.
NLP, classification and LLM evaluation
Intent, topic, sentiment, toxicity and entity labeling, plus scoring of LLM responses for chatbots, assistants and internal tools.
Document labeling and retrieval
Segment and tag documents, define relevance labels and build ground-truth sets for testing retrieval and RAG answers.
Computer vision and content safety
Image categories, bounding boxes, content-safety tags and media classification for risk, compliance or user experience.
Structured data tagging
Events, anomalies and custom categories across tables, logs and monitoring data, for ML and analytics.
Development and public-sector projects
Labeling for governance, infrastructure monitoring, survey data and donor-funded programmes, where consistency gets checked.
Domain-specific labeling
We work with your subject-matter experts on task-specific taxonomies and examples for your sector and risk profile.
Quality, security and tooling
We run labeling with the same documented processes, records and support practices as our software work.
Quality controls
Security and privacy
Tools and integration
We can work in your labeling platform or set one up for you, and we coordinate with your data engineering and MLOps teams on formats and hand-offs.
Data labeling: frequently asked questions
What types of data can you label?
Text, documents, images, short video and audio segments, and structured or tabular data. For a custom format we usually run a pilot and some extra onboarding first.
Can you evaluate LLM outputs and check them for safety?
Yes. We label LLM outputs for quality, relevance, safety and policy compliance, and run side-by-side comparisons of prompts or models so you can choose a configuration on evidence.
Will you work inside our labeling tool?
Yes, if you have one. Otherwise we recommend and configure a third-party or custom tool, depending on your security, volume and integration needs.
How do we get started?
A short discovery call about your data, use case and success criteria, then a written proposal within two working days for a pilot batch and guideline design. We scale up after you accept the pilot.
How do you know the labels are right?
You measure it. We mix gold tasks with known answers into the work, track how often labelers agree with each other, and review in layers before anything is delivered. Every batch ships with a report of those numbers, so the dataset has a paper trail when your model, or an auditor, is questioned later.
| In every batch report | What it tells you |
|---|---|
| Inter-annotator agreement | How often two labelers give the same answer. Low agreement means the guideline is unclear, not that the labelers are careless. |
| Gold-task accuracy | How labelers score on items with a known answer, mixed into the normal work. |
| Error types | Which labels get confused with which, so the guideline can be fixed at the source. |
| Rework | How much was corrected at peer or lead review before delivery. |
| Volume and turnaround | Items done against plan, and anything blocked waiting for a decision from you. |
From first call to first batch
The pilot comes before any volume commitment. If the guidelines need a second round, that happens on a small batch, not on your whole dataset.
After the first call you get a written proposal within two working days.
Discovery call
Thirty minutes, free. Your data, the task, the volume, the deadline and how the labels will be used.
Written proposal (within 2 working days)
Pilot scope, quality targets, tooling, data-handling terms and pricing.
Guidelines and pilot batch
Worked examples, edge cases and a first batch. We measure agreement and fix the guideline where labelers disagree.
Production with layered QA
Self-check, peer review and lead review. Questions about edge cases come back to you rather than being guessed.
Batch reports and handover
Each delivery comes with the quality report. Guidelines are updated as your model's needs change.
Data stays under your controls
We can label inside your own tool and environment, with access you grant per person and remove when the work ends. NDAs are signed before any data is shared.
A documented dataset
Guideline versions, batch reports and review records are kept under our ISO/IEC 20000-1 certified processes, which helps when governance frameworks such as the EU AI Act or NIST AI RMF ask how training data was produced.
Connected to the rest of the pipeline
Labels arrive in the format your training or evaluation code expects. For the pipeline itself, see data engineering and AI/ML engineering.
Which engagement fits?
| Option | Use it when | What's fixed up front |
|---|---|---|
| Pilot batch | New use case, new data source, or a first project with us | Guidelines, batch size and quality targets |
| Ongoing monthly team | Data keeps arriving and models are retrained regularly | Team size and monthly volume, on 30 days' notice |
| Project-based labeling | One dataset or evaluation campaign with a deadline | Scope, milestones and acceptance criteria |
| Labeling specialists in your team | You run your own labeling operation and need more trained people | Roles and start date, usually 5–10 working days |
Pricing is scoped and quoted after a free 30-minute discovery call, with a written proposal within two working days.
When this isn't the right fit
- You need millions of very simple labels and the lowest possible cost per item. A crowd platform will be cheaper; we focus on work where guidelines and review matter.
- The labels need a licensed professional, such as a clinician, and you can't provide one to guide and review. We won't fake domain expertise.
- You need a round-the-clock labeling queue. Our team works 09:00–18:00 Nepal Time.
DAIR, the Dignep AI Research Unit
Related research on language data:
DAIR is our applied AI research unit, led by Dr. Yagya Raj Pandeya of Kathmandu University's Department of Artificial Intelligence. Its work covers low-resource NLP for Nepali, retrieval-augmented generation, and safety tooling such as hallucination detection and red-teaming.
Further reading
Start with a pilot batch
Book a free 30-minute call. You get a written proposal for a pilot within two working days.
