Skip to content
NewRL environments

Expert human data for frontier AI.

Data annotation, RLHF and RL environments. A second expert checks every task.

For enterprisesBecome an expert
Data annotationRLHFRL environmentsData licensing

Services

Four services. One team.

Labels

Data annotation

Text, code, tables and audio, labelled by domain experts.

  • Classification
  • Entity labels
  • Transcription
  • Gold sets
Get data labelled
Your data

Data licensing

We prepare your records and license them to AI labs.

  • Code records
  • Support tickets
  • Rights checked
  • Source tracked
License your data
New

RL environments

Agent tasks with a grader and a correct answer.

  • Agent tasks
  • Graders
  • Reward rubrics
  • Trajectory review
How it works
Human feedback

RLHF and evaluation

We rank model outputs, find errors and write correct answers.

  • Preference ranking
  • SFT examples
  • Red-teaming
  • Eval sets
Start a pilot

RLaaS

RL environments that agents learn from.

We build the task, the grader and the reward rubric.

Real
Real questions, databases and code.
Checked
A reference answer and a grader for each task.
Updated
Harder tasks when your model improves.
Talk to us about RL
The RL loop
  1. Task01

    An expert writes the task.

  2. Agent02

    Your model tries to solve it.

  3. Grader03

    Code and an expert score it.

    Expert scores here
  4. Reward04

    The score goes back into training.

env = Task(db="retail_q3", tools=[run_sql, chart])
traj = agent.rollout(env)
score = grader(traj) + expert.score(traj, rubric)
→ reward=0.62 · error="refunds not netted"

How it works

You get accepted work, not resumes.

We test, train, manage and replace the experts. You work with one team.

  1. 01

    Scope

    Agree the task and the quality bar.

  2. 02

    Test

    Experts pass a test on your task.

  3. 03

    Train

    Experts practise on your examples.

  4. 04

    Deliver

    Work runs on your tools or ours.

  5. 05

    Scale

    Weekly reports. You choose when to grow.

Our experts

Practitioners, not a crowd.

Learners and graduates in tech, data, product and marketing. Tested on your task before they start.

Skill area

Python and SQL agent tasks

PythonSQLJSONFunction calls
500Profiles with these skills in our sample.
Where our experts are
  • Maharashtra
  • Karnataka
  • Uttar Pradesh
  • Delhi
  • Telangana

Counts come from a sample of resumes. Skill areas overlap.

Quality

You set the quality bar.

We agree the targets before we start. You get a report each week.

  • First-pass acceptance
  • Critical errors
  • Reviewer agreement
  • Turnaround
  • Rework
The pilot

Start with 20.
Scale to 100.

people in the first batch
20
people in the first batch
pilot length
2 weeks
pilot length
experts per domain
~100
experts per domain
Start a pilot

Your data

License your data to AI labs.

Records with real errors and fixes are valuable. We check the rights, prepare the data and manage the licence.

We share nothing until the rights are clear.

License your data

  • Code records

    1. Attempt
    2. Failed tests
    3. Feedback
    4. Fix

  • Support tickets

    1. Ticket
    2. Policy
    3. Action
    4. Resolved

  • Internal workflows

    1. Request
    2. Tool steps
    3. Approval
    4. Outcome

FAQ

Common questions.

Who are the experts?

Learners and graduates from our network in India. Each expert passes a test before they start.

Do you use our tools?

Yes, if you want. If you have no tools, we use an existing annotation tool.

What does a pilot cost?

It depends on the task, the volume and the quality bar. You get the full cost breakdown.

Can you license our data?

Yes. We check your rights first. Then we clean, label and license the data to AI labs.

How do you protect data?

We clear permissions, IP and privacy terms first. We never reuse client work without written permission.

Get in touch

For enterprises

License your data or get it labelled.

Enterprise enquiry
For experts

Paid annotation work on fair terms.

Apply as an expert

Or email hello@neevlabs.ai