கKatralAI
Nila

Nila

Campus mentor

Foundations, studio projects, then a fork — builder, advisor, or domain.

Constellation

Campus Katral

Foundations, studio projects, then a fork — builder, advisor, or domain.

0/18 lessons · 6 modules

Next on this path: Words are not how AI sees words

You leave with

A campus project you can walk through

One handbook, twenty checks, a refuse bell, a three-minute walkthrough. You can teach it.

You will be able to

  • Explain tokens, retrieval, and tool loops without mythology.
  • Choose retrieval for living facts and fine-tune for behaviour.
  • Ship a narrow project with checks, a refuse path, and a walkthrough.

Module 1 · Language machines

  1. Lesson 1 · 10 min · 55 XP

    Words are not how AI sees words

    A language model predicts the next token. A token is a shard of text — not a thought.

  2. Lesson 2 · 12 min · 60 XP

    Closed book vs open notes

    RAG is retrieval-augmented generation: give the model the right page before it answers.

Module 2 · Desk, tools, loops

  1. Lesson 3 · 10 min · 50 XP

    The desk has a size

    The context window is everything the model can see at once. Overflow, and early notes fall off the desk.

  2. Lesson 4 · 11 min · 55 XP

    A tiny campus agent

    An agent is a loop with tools: search notes, draft, check. Your project is a campus FAQ bot that refuses guesses.

Module 3 · Integrity and proof

  1. Lesson 5 · 9 min · 50 XP

    Integrity is the grade

    Using AI to think is a skill. Submitting AI as your mind is a fail. Learn the difference.

  2. Lesson 6 · 11 min · 55 XP

    Measure before you brag

    A campus demo is not a grade. An eval set is how you prove the model did the job.

Module 4 · Building blocks

  1. Lesson 7 · 10 min · 55 XP

    Neighbours in meaning-space

    Embeddings put text on a map. Nearby points mean similar ideas — that is how search finds the right page.

  2. Lesson 8 · 10 min · 55 XP

    Prompts that hold

    A prompt is a spec: role, goal, constraints, examples, output shape. Vague specs make fluent junk.

  3. Lesson 9 · 10 min · 55 XP

    Tools, not magic

    Tool use is a loop: choose a tool, read the result, decide. The model is not browsing the universe — it is calling functions you gave it.

Module 5 · A project that hires

  1. Lesson 10 · 10 min · 55 XP

    Pictures, voice, text

    Multimodal models braid image, audio, and text. They still guess. A photo is not a proof.

  2. Lesson 11 · 10 min · 55 XP

    Fine-tune vs RAG

    Fine-tune to change behaviour. RAG to change facts. Mixing them up is how campus demos lie.

  3. Lesson 12 · 14 min · 80 XP

    A campus project that hires

    Certificates are receipts. A tiny RAG over a real handbook, with evals and a refuse path, is proof.

Module 6 · Production craft

  1. Lesson 13 · 12 min · 70 XP

    Temperature is a dice, not a mood

    At heat 0 the model repeats the same safe sentence. At high heat it writes a pretty lie. Facts like cool dice.

  2. Lesson 14 · 12 min · 70 XP

    Cut by meaning, not by length

    Equal strips mix two topics and poison retrieval. A chunk is one idea you can cite.

  3. Lesson 15 · 12 min · 70 XP

    Retrieve wide, then rank

    First search is cheap and noisy. A second pass keeps the three strips that actually answer.

  4. Lesson 16 · 12 min · 70 XP

    The user is not the boss of the rule

    Ignore previous instructions is a trick, not a password. The hidden rule still wins.

  5. Lesson 17 · 12 min · 70 XP

    Ask for a form, not a poem

    A gate cannot read fog. Hours, room, person must be fields — or the next tool will invent them.

  6. Lesson 18 · 12 min · 70 XP

    A gold thread through the miss

    If you cannot replay the question, the chunks, the prompt, and the gate, you cannot fix a silent drop.