Skip to content

Seven kinds of work.
One standard.

Our people do this work every day. Teams can use it two ways: put our people on your project, or hand us the spec and get reviewed data back.

01

Data labeling

Boxes, polygons, spans, intents, sentiment, toxicity. Every guideline gets read, argued over in calibration, then applied the same way on task one and task ten thousand.

  • Bounding boxes & segmentation
  • Span & entity tagging
  • Classification & moderation
  • Video event marking
02

Preference & RLHF

A preference without a rationale is noise. We rank, score against rubrics, and write the why — so your reward model learns the judgement, not the coin flip.

  • Pairwise & listwise ranking
  • Rubric scoring
  • Written rationales
  • Response rewrites
03

SFT & expert writing

Natural prompts, multi-turn conversations, and reference answers from people who know the subject — from everyday chat to STEM reasoning and long-form instruction following.

  • Prompt authoring
  • Gold responses
  • Multi-turn dialogue
  • Reasoning traces
04

Audio & video

Recorded speech in real rooms and real accents, verbatim transcription, speaker turns, captions, and frame-level review of video datasets.

  • Speech collection
  • Verbatim transcription
  • Diarization
  • Captioning & video QC
05

Indian languages

Native speakers, not translators reading a glossary. We write, translate, localise and judge model output in Telugu and Hindi, and add languages as the team grows.

  • Native prompt writing
  • Translation & localisation
  • Fluency judging
  • Code-mixed text
06

Evals & red-teaming

We design the rubric, find where the model breaks, label the failure mode, and grade without grade inflation. Useful when you need to know, not hope.

  • Rubric design
  • Adversarial prompting
  • Failure analysis
  • Benchmark grading
07

Agents & code

Terminal-bench style tasks, coding trajectories, tool-use traces and agent runs — written, executed and judged by people who ship software for a living.

  • Terminal tasks
  • Trajectory review
  • Unit-test authoring
  • Tool-use grading

Project families
we've worked.

Internal codenames from programs our people have contributed to. If yours is here, we already know the guidelines.

  • Riga
  • Otter
  • Geranium
  • Gauss
  • Terminus 2.0
  • Tuna
  • Marlin v3
  • Sentinel Ultra
  • Draco
  • Stoa
  • Octahedron
  • Hertz
  • Cipher
  • Generalizt
  • Dynamo
  • and more

Three ways in.

Start here

Pilot

A small, fully reviewed batch on your guidelines. You score it. No long contract.

Most common

Dedicated team

A trained, managed team that works only on your project, with a lead and QA built in.

Least overhead

Managed delivery

Send the spec and an acceptance bar. We return reviewed data, and you only see finished work.