Learn how people make AI better.
Start with a short lesson. See a real example, make a judgment, and create something you can explain in your own words.
6learning paths
18modules
EN + हिंदीin every lesson
Every lesson follows the same four steps
- LearnOne idea in plain words.
- SeeA weak and a stronger example.
- TryA small decision with feedback.
- MakeYour own private practice sample.
What you can learn
Curriculum reviewed:
AI trainer essentials
Practise the core judgment, ranking, explanation, and rewriting work used in human-feedback projects.
- Read the instruction and rubricTurn a broad request into observable checks before judging an answer.18 minutesNew learner
- Compare two model responsesChoose the better response by applying the rubric, not by length, confidence, or writing style alone.22 minutesNew learner
- Explain and improve a judgmentLocate the specific failure, write a concise rationale, and choose the revision that fixes it.24 minutesSome basics
Bharat language & voice
Evaluate naturalness, meaning, code-switching, speech, and local context without flattening language differences.
- Judge meaning and naturalnessSeparate literal correctness from wording a fluent speaker would actually use.20 minutesNew learner
- Review code-switching and contextJudge whether mixed-language wording matches the audience, register, and meaning instead of penalising it automatically.18 minutesNew learner
- Evaluate speech and transcriptsDistinguish what was spoken, what is audible, and what a transcriber has inferred.22 minutesSome basics
Expert, safety & agents
Review factual claims, domain boundaries, harmful failures, and the tool traces produced by AI agents.
- Check claims and evidenceSplit an answer into checkable claims and distinguish evidence from plausible wording.24 minutesSome basics
- Classify and escalate safety failuresDocument a reproducible failure without amplifying harmful content or operating outside the evaluation boundary.25 minutesSome basics
- Review an agent traceFind the first unsupported or unsafe tool action, then propose a recoverable next step.24 minutesComfortable with the basics
How post-training works
Understand where demonstrations, preferences, reward signals, and evaluation fit. No coding is required.
- Demonstrations and SFTSee how carefully written examples teach a model the desired form of an answer.18 minutesNew learner
- From preferences to a training signalUnderstand what pairwise choices capture, and what they can miss when guidelines or reviewers disagree.20 minutesSome basics
- Evaluation closes the loopUse held-out tasks and explicit criteria to check whether a change helped without creating new failures.20 minutesSome basics
Multimodal & physical AI
Practise image, document, video, and robot-action review with careful grounding and uncertainty.
- Ground answers in image and videoDescribe only what the media supports, identify the relevant region or moment, and flag ambiguity.24 minutesNew learner
- Review documents and OCRCheck transcription, reading order, tables, and unclear fields without silently inventing text.22 minutesNew learner
- Review robot-action tracesSegment an action into observable steps and locate grasp, contact, motion, and safety failures.28 minutesSome basics
Coding & reasoning evaluation
Evaluate code, maths, science, and research answers by testing claims and locating the first invalid step.
- Test and review generated codeUse requirements and tests to find the first correctness, edge-case, or safety failure.26 minutesComfortable with the basics
- Find the first invalid reasoning stepCheck assumptions, units, calculations, and conclusions in order instead of judging only the final answer.25 minutesSome basics
- Verify claims and citationsSplit an answer into checkable claims, prefer primary evidence, and state what remains uncertain.24 minutesSome basics