ishu garg

AI Labbuild / test / explain

AI that survives contact with users, documents, and bad assumptions.

The path here is practical: build small, evaluate early, keep failure diaries, and do not confuse a confident demo with a working system.

build / eval / ship

A lab for building useful AI systems without mistaking a demo for proof.

Start with small shipped artifacts: prompts as interfaces, RAG with source boundaries, evals before confidence, and agents with visible limits.

Start zero to shipped agent Browse AI archive
Featured
  1. LLM Builder: Zero to Shipped Agent Tracks / beginner / 6 weeks Active
  2. Agent Failure Zoo Essays / beginner / series seed Active
  3. Break The LLM Challenge Resources / beginner / challenge shelf Seed
  4. BhumiSahayak Projects / deep / active Active
  5. AI Student Helper System Projects / intermediate / project seed Seed
tracks

Build from zero to artifact

Structured paths for LLMs, agents, from-scratch models, and eval-heavy build habits.

Growth rule

Tracks get promoted when they end in a concrete project and checkpoint list.

failure

Break toy systems safely

Prompt injection, hallucination, eval gaps, bad citations, and tool-use boundaries.

Growth rule

Failure cases need a reproduction, prevention pattern, and regression test idea.

proof

Projects with evidence

Public-safe project notes and applied-AI systems with boundaries visible.

Growth rule

A project page must say problem, boundary, evidence, and reusable lesson.

source

Sources, app shells, and deployment boundaries

Radar packets, RAG sources, AI app hosting, static shells, and security/performance posture.

Growth rule

Source and app-boundary notes need a static public shell, current source links, and explicit security posture.

Build queue
  • RAG source-boundary checklist - source dossier / next
  • Tiny attention-head visualizer - interactive widget / interactive-later
  • AI app hosting map - tooling guide / seeded
Build graph

From idea to shipped agent.

The AI room is a process graph: frame the job, build the interface, retrieve evidence, evaluate failures, add tools, then ship with limits visible.

  1. 01 Frame Pick one user, one job, one failure that would matter. problem note
  2. 02 Prompt Treat prompting as interface design, not wishful wording. prompt contract
  3. 03 Retrieve Give the model sources, boundaries, and stale-data rules. source policy
  4. 04 Evaluate Write pass/fail cases before demo confidence arrives. eval set
  5. 05 Agent Add tools only where exact action or fresh state is needed. tool boundary
  6. 06 Ship Log failures, explain limits, and keep humans in the right loop. release note
Open related archive
Structure

AI Lab shelves.

What can I build, test, explain, or break safely?

build path

Zero to shipped

Tracks and ladders for building LLM systems, tiny models, and useful agent tools.

Route-Tabbed Subtopic Hub 4 objects

Growth rule

Show path rows first; full module detail belongs on the track route.

View all / route tabs plus shelves

failure zoo

Break before believing

Prompt injection, tokenizer traps, hallucination checks, and eval cases.

Challenge Board 8 objects

Growth rule

Surface one trap, one eval object, and one safe sandbox link.

View all / difficulty/technique lanes

evals

Tools, RAG, and evidence

Reusable patterns for retrieval, sources, evals, logs, and bounded tool use.

Source Table 11 objects

Growth rule

Facts and framework claims need sources; demos need failure modes.

View all / table or row matrix

watchlist

Exciting tech worth tracking

AI, robotics, interfaces, biotech, and space ideas that may become essays or demos.

Curated Shelf Plus Archive 1 object

Growth rule

Keep watchlist entries short until sources and owner views are ready.

View all / editor picks plus full archive link

route anchors

Other objects on this route

Compact anchors preserve old links without turning this page into a card wall.

More anchors
How this room grows
  1. site Whole site Keep the homepage as orientation and taste. Move depth to rooms and the archive.
  2. area Room / area Create a route when the audience, mental model, or review rules differ from siblings.
  3. subtopic Shelf inside a room Promote to its own page when it needs custom interaction, more than 20 objects, or a recurring newsletter/search path.
  4. collection Curated slice Use rows for many similar objects and cards only for heterogeneous featured objects.
  5. object Canonical content object Every public object keeps a detail URL and machine-readable summary, even if surfaced in many rooms.
  6. block Reusable block Blocks can appear in many posts. They never own taxonomy color or public identity by themselves.
Open full archive
Tracks

Builder tracks as module ladders.

beginner / 6 weeks

LLM Builder: Zero to Shipped Agent

Ship one useful AI assistant with a clear task, test set, retrieval boundary, and failure log.

  1. 01 Prompting as interface design Turn a messy note into a structured student resource.
  2. 02 RAG without superstition Build a scholarship finder over a small verified corpus.
  3. 03 Evals before vibes Evaluate the scholarship finder on 20 realistic questions.
  4. 04 Tools and agents Add a deadline checker and source verifier.
  5. 05 Ship and observe Publish a read-only demo with a failure diary.
intermediate / 8 weeks

LLM From Scratch, No Incense Required

Train a tiny text model, inspect its failures, and connect the toy version to modern LLM systems.

  1. 01 Text to numbers Write a tokenizer visualizer for your own bio.
  2. 02 Embeddings Plot nearest neighbors in a toy corpus.
  3. 03 Attention Animate one attention head on a short sentence.
  4. 04 Training loop Train a tiny quote generator.
  5. 05 Failure diary Publish the model's wrong outputs with explanations.