Login AI
AI data labeling
We label datasets for model training and evaluation: we write the guidelines and taxonomy, label four modalities, resolve annotator answers through consensus, and hand over a QA report per class.
We prepare labeled text, images, audio, and video against agreed rules with two review passes.
Pricing
from ₽300,000
per project
This is the lower bound of a project. We scope the exact quote after a brief: it depends on volume and the number of integrations.
Capabilities
Guidelines and taxonomy
We build the class tree, definitions, and borderline examples. Guidelines fix what counts as a label and cut the spread between annotators.
Four modalities
Text, images, audio, and video run through one pipeline: each modality gets its own tool, shared acceptance rules, and a decision log.
QA and consensus
Every tenth item gets a second pass. We resolve disagreements into consensus and compute agreement per class, then keep it in the report.
RLHF and preference pairs
We collect answer pairs for RLHF: an annotator picks the better option against a rubric, and disputed pairs go to an expert for arbitration.
From request to result
1
Taxonomy audit
We review the current classes, examples, and borderline cases. We check that each class has a definition and an observable signal.
2
Guidelines and pilot batch
We write the instruction with examples and label a pilot of 200–500 items. We measure agreement and fix ambiguous spots before the stream starts.
3
Labeling pipeline
We set up queues, roles, and acceptance. Batches run on schedule, progress shows in a dashboard, and each label keeps its author and version.
4
QA and handover
We run the sample through two passes, resolve disagreements, and assemble the report. We hand over the dataset in the agreed format along with the guidelines.
Check if it fits you
Good fit
+ 10,000 items
A volume of 10,000 items or more pays for the guidelines and pipeline setup: batches run as a stream, and quality rests on the sample.
A domain expert on hand
An expert sets the labeling reference: the other annotators and QA use their decisions.
Not a fit
No taxonomy
Without a class tree and definitions, labeling drifts: the audit and guidelines come first, then the batches.
PII without a policy
Personal data needs a retention and anonymization policy: until you agree on it, batches with PII stay on hold.
What you get
10,000
10,000 items in the base
The base batch covers 10,000 items and gives enough examples for training and a measurement on a held-out split.
text
Text as the base modality
The text modality goes first: text acceptance takes one QA pass, and the taxonomy template carries over to images, audio, and video.
2
Two QA passes
Two QA passes give agreement per class and show which definitions are worth tightening before scaling.
What the delivery includes
Guidelines and taxonomy
We hand over the taxonomy, class definitions, borderline examples, and an acceptance rubric. The document records its version and is the reference for annotators.
A labeled pilot batch
We deliver 200–500 items with labels, versions, and a decision log. The pilot batch shows agreement on real data before scaling.
A QA report
We assemble a report on agreement, disagreements, and class coverage. The QA report shows where the rules hold and which classes need more work.
How the system is built
Data intake
We accept files in batches, check the format and completeness, route them into queues, and assign an owner to each batch. At intake we filter broken files and duplicates, so the annotator receives a clean set.
Labeling
Annotators label against the guidelines and rubric, leave a comment on a borderline case, and pass the item to a second pass. Each label keeps its author, time, and instruction version, so the decision history can be restored.
QA and consensus
A second annotator reviews the sample, disagreements go to expert arbitration, and the final label is fixed through consensus. We compute the agreement metric per class and return disputed definitions to the guidelines.
Where the solution pushes back
- 01
The quality of the guidelines
The quality of the guidelines sets the label spread: precise definitions and examples raise agreement and cut fixes at acceptance.
- 02
Expert availability
Expert availability sets the arbitration speed: borderline cases wait for the expert decision, so the expert schedule sits in the batch plan.
- 03
The PII policy
The PII policy defines the workflow: we agree on anonymization and retention before the start, and that adds a step to batch preparation.
Technologies
The technologies we use in our work and their role in a project.
Labeling pipeline
We write validation and metric scripts in Python, store tasks and statuses in PostgreSQL, keep dataset files in MinIO, run the job queue through Kafka, and hold batch progress in Redis.
Cost calculator
This is an indicative range, not a quote. We scope the exact estimate after a brief: it depends on volume and the number of integrations.
Items to label
Modality
Quality level
Guidelines
Add-ons
Support
Year one: updates and monitoring, a reply within 2 business days.
Service
AI data labeling
The estimate covers our team's work. Hardware, third-party service licences and travel are added on request.
Estimated cost
₽400,000
Timeline
2 wk
Frequently asked questions
How do you control labeling quality?
Every tenth item gets a second pass, we resolve disagreements into consensus, and we compute agreement per class. We fix the acceptance threshold in the guidelines before the batch starts.
Can you label under an NDA and on our own infrastructure?
Yes. We sign an NDA, work in your cloud or on an on-prem deployment, grant file access by role, and keep the action log on your side.
Which modalities do you label?
Text, images, audio, and video. Each modality has its own tool and acceptance rules, while the taxonomy and QA stay shared.
How long does a pilot take?
The pilot runs 1–2 weeks: a taxonomy audit, the guidelines, and a batch of 200–500 items with an agreement measurement. After the pilot we quote the timeline for the full volume.
What teams add to this service
Related Services
Part of Solutions
Implementation Examples
−90%
Escaped defects
Ready to discuss your task?
Leave a request – we'll get back with a proposal and a preliminary estimate.
Contact us