NEW Evaluation & Observability: all 5 parts are live
Part 1 of every course is free to read.
Every brief is free to read.
Roadmaps
Twenty questions of every kit are free.
A free 20-minute call first.
Start reading free
The AI engineering ecosystem

Learn. Build. Ship.

Courses that explain AI engineering properly. Production projects to build the real thing. Real interview questions to prove it stuck. Everything in one place.

CoursesProduction projectsInterview kits1:1 sessions

A free concept every weekdayCourses bought once, never rentedNothing renews, ever

Most read this week Needs real data

Transformers, Illustrated

Cover of Transformers, Illustrated
Parts7, and part 1 is free to read in full Reading3h 28m, measured not estimated Drawn34 diagrams, made for this course Price₹2,400 once. Nothing renews.
View course →
Just published 21 August

Evaluation & Observability is complete. All five parts are live, and they are added to the course for everyone who already owns it. No extra cost.

That is how every course works here. You buy it once, and it keeps growing.

See what changed →
Build it yourself Brief · free to read

Build attention from scratch, then check it

  • A working single-head block, in numpy only
  • Matching a reference implementation within tolerance
  • No step-by-step. The brief withholds it on purpose.
Read the brief →
Interview kit 20 of 640 free

“A model handles 4k context fine and falls over at 32k. Walk me through where the cost actually goes.”

What is being probed: whether you know attention is quadratic in sequence length and the KV cache is linear in it, and which one bites first.

See a strong answer →
Career services Paid, per engagement

Someone moved roles this month

FromData engineer → ToAI engineer

Four mock interviews, a rewritten profile, and eight weeks. We will name them here once they say yes. Placeholder · needs a real, consenting person

How the help works →
Today · free to read No account · 9 min

What attention heads actually learn

A head is a question the model asks about a sentence, and the answer is a weighted average of what it already knows. Take a concrete one. “The keys that the engineer dropped were hers.”

To resolve hers, something has to reach back to engineer, skipping keys, which is nearer and grammatically tempting.

Keep reading · 8 min left →

Recently published

The latest writing, newest first.

Everything published →

A subject, in order,
from first idea to working code.

Every course opens with a free first part.

1 course
Part 1, free, no account

Python

for AI Engineers

View course
79 parts · 9h 13m
1
All courses
Browse the full catalogue, with filters and search.
Browse all courses →
The idea here: what attention heads actually learn

One idea, four depths.

Every row below is that same idea, at a different level of commitment. Read it in a minute, follow it through a course, build something that cannot work without understanding it, or sit across from someone asking you about it.

  1. 01 Concept Free, no account

    What attention heads actually learn

    A handful of heads copy the token that followed the last time this one appeared. That is the whole trick.

    1 min to skim · 9 min in full
    Read it →
  2. 02 Course ₹2,400 once

    Transformers, Illustrated

    1. Tokenisation and embeddings
    2. Attention from first principles
    3. What attention heads actually learn
    4. Induction heads
    5. 3 more parts
    7 parts · part 1 free · later parts included
    See the course →
  3. 03 Project Ships with its course

    Find the induction heads in a small model

    A brief, not a walkthrough. Finding them needs the architecture, an evaluation you trust, and enough Python to instrument a forward pass.

    1 weekend · a repo, a write-up and one number
    See the projects →
  4. 04 Interview kit 20 free

    Could you answer these?

    • Why does a head specialise at all?
    • How would you find an induction head?
    • What breaks when you scale heads?
    24 questions on this subject alone
    Try one →

Every concept is free.
So is the deep dive behind it.

Every idea gets the one-minute version. Most also have a deep dive: the same idea followed all the way down, ten to twenty minutes, with the code that runs and the part where it breaks. Both free.

Posted daily on Instagram. Kept here so you can find it again
Security & safety

Prompt injection

Anything your model reads can give it instructions. A web page, a PDF, a support ticket. If it lands in the context window, it competes with your system prompt.

01Untrusted text enters the context: retrieved doc, tool output, user paste.
02The model cannot tell your instructions from the text's instructions.
03It follows whichever is more specific. Usually the attacker's.
Read the deep dive · 14 min → Free, no account
Part of Security & Safety · 9 concepts

Browse the concepts

Six areas, one minute each. Pick one to read it above.

Security & safetyPrompt injectionAnything your model reads can give it instructions.1 mindeep diveSecurity & safetyJailbreaksPrompts crafted to talk a model out of its own rules, and why filters alone do not hold..1 mindeep diveSecurity & safetyData exfiltration via toolsA tool call is an open door out.2 mindeep diveSecurity & safetyPII leakageModels repeat what they were shown.1 minSecurity & safetyModel supply chainWeights, datasets and adapters you did not build are dependencies, with everything that implies..2 minSecurity & safetyOutput validationNever act on raw model output.1 minSecurity & safetySandboxing tool callsA model that can run code needs walls: least privilege, timeouts, and no network by default..2 minSecurity & safetyRate-limit abuseYour endpoint is someone else's free compute until quotas, keys and budgets say otherwise..1 minSecurity & safetyRed-teaming a promptAttack your own system prompt before someone else does.2 mindeep diveRetrieval & RAGChunkingHow you split documents decides what retrieval can ever find.1 mindeep diveRetrieval & RAGEmbeddingsText as points in space, where distance means similarity.1 mindeep diveRetrieval & RAGHybrid searchVectors miss exact terms; keywords miss paraphrase.2 mindeep diveRetrieval & RAGRe-rankingA second, slower model reorders the shortlist so the best passage is actually first..1 mindeep diveRetrieval & RAGChain of retrievalOne search is rarely enough.2 mindeep diveRetrieval & RAGContext stuffingMore passages is not more signal.1 minRetrieval & RAGStale index driftYour corpus changed and your index did not.1 minRetrieval & RAGCitations that hold upMaking a model point at its sources in a way you can actually verify..2 minArchitecturesAttentionEvery token looks at every other token and decides what matters.1 mindeep diveArchitecturesInduction headsThe attention heads that learn to copy patterns, and why in-context learning works at all..2 mindeep diveArchitecturesKV cacheWhy generation gets cheaper after the first token, and what that costs in memory..1 mindeep diveArchitecturesMixture of expertsRoute each token to a few specialists instead of the whole network.2 minArchitecturesPositional encodingAttention has no idea of order.1 minArchitecturesContext window limitsWhy long context is hard, what degrades first, and what the marketing numbers hide..1 mindeep diveArchitecturesTokenisationModels do not read words.1 minTraining & tuningLoRAFine-tune a giant model by training two thin matrices beside it.1 mindeep diveTraining & tuningQuantisationFewer bits per weight, smaller and faster models, and the accuracy you trade for it..1 mindeep diveTraining & tuningRLHFTraining a model on which answers people prefer, and what that does to honesty..2 mindeep diveTraining & tuningCatastrophic forgettingTeach a model something new and watch it quietly lose something old..1 minTraining & tuningDistillationA small model trained to imitate a large one.1 minTraining & tuningSynthetic dataTraining on model output: when it compounds quality and when it compounds error..2 minTraining & tuningWhen not to fine-tuneMost fine-tuning tickets are retrieval problems in disguise.1 mindeep diveEvaluationHallucination rateYou cannot fix what you do not measure.1 mindeep diveEvaluationLLM as judgeUsing a model to grade a model: cheap, scalable, and biased in ways you must control..2 mindeep diveEvaluationGolden setsA small set of answers you trust completely, and everything you can anchor to it..1 mindeep diveEvaluationOffline vs online evalThe benchmark says it improved.2 minEvaluationDriftThe inputs change under you.1 minEvaluationRegression suitesEvery prompt change can break something that worked.1 minServing & costBatchingServing many requests in one forward pass.1 mindeep diveServing & costSpeculative decodingA small model drafts, the big model checks.2 mindeep diveServing & costToken economicsPrompt tokens and completion tokens price differently.1 mindeep diveServing & costLatency budgetsDecide what slow means before users do, then spend the budget deliberately..1 mindeep diveServing & costCaching responsesCache at the semantic layer, not the string layer.1 minServing & costCold startsThe first request after idle pays for everything.1 minServing & costStreamingTokens as they generate.1 minServing & costCost per resolved ticketThe only cost metric that survives contact with a CFO.2 min
45 of 312 concepts in the library. One more every weekday. Search anything with ⌘K. Browse all concepts
Projects

A concept you can explain
is not yet a thing you can show.

A course teaches one subject. A project refuses to: it hands you a problem that needs three or four of them at the same time, which is what the job is. Do one before you open an interview kit. Most of what a kit asks is about work exactly like this.

  1. It cuts across topics.A course follows one subject all the way down. A project needs three or four of them at once, which is how the work actually arrives.
  2. A brief, not a tutorial.You get the problem, the data and the review checklist. There is no walkthrough to copy, because copying one proves nothing.
  3. An artefact, in a weekend or two.A repository, a short write-up and one number you can defend. Scoped so you finish it, because an unfinished project proves nothing.
Project briefs ship with their course. The dataset, the review checklist and the write-up template come with each one. See the project briefs
Interview kits

2,240 questions,
from real interviews.

A kit is every question we have collected for one role, grouped by the topic it tests and tied to the course that teaches it. Pick your role to see what is in yours.

640 questions in the AI Engineer kit

20 from every kit are free. The rest come with the course each topic is drawn from.

  • 240questions38% of this kitRetrieval & RAGFull set with RAG Systems in Production
  • 180questions28% of this kitSystem design for LLM appsFull set with Serving & Scaling
  • 120questions19% of this kitEvaluation & observabilityFull set with Evaluation & Observability
  • 100questions16% of this kitCost & latencyFull set with Cost & Latency Engineering

Try one from this kit

Four sample questions. Say honestly whether you could answer each one.

Your readiness · AI Engineer
0 / 4
Answer the questions to see where you stand.
The question
RETRIEVAL · ASKED AT 6 COMPANIES
Your RAG system returns confident, wrong answers. Walk me through how you would find out why.
Senior AI Engineer · 45-minute system design round
What a strong answer covers included with every question
  • →Separate retrieval failure from generation failure before touching either
  • →Measure recall@k on a golden set rather than eyeballing outputs
  • →Check chunk boundaries. The answer often spans two chunks
  • →Name the four failure modes and which metric surfaces each
Could you answer this in an interview?
  • Every question carries its context.The role, the round and the seniority it was asked at, so you know what level the answer has to reach.
  • Every question carries an answer outline.Not a scripted answer. The points a strong one has to hit, which is what an interviewer is actually listening for.
  • 20 per kit are free.No account. The rest come with the course that teaches that topic, so a kit is never sold on its own.
20 questions from every kit are free. The full kit comes with its course. Browse the kits

Testimonials

What engineers
say.

Some reviewers asked to stay anonymous while they were still interviewing.

4.8average across 4 reviews

  • The AI Engineering Kit gave me the comprehensive end-to-end knowledge I needed. It perfectly bridged the gap between my software background and modern AI production needs.
    SrivaniSDE @ Infosys
  • The LLM evaluation and observability content is genuinely production-grade. Way more practical than any course I'd tried before.
    AnonymousML EngineerIdentity protected
  • PracticAI is the reason I stopped freezing up in system design rounds. The production project walkthroughs made abstract MLOps concepts finally click.
    AnonymousData Scientist → AI EngineerIdentity protected
  • I especially liked the sections on hybrid search, evaluation, and ingestion pipelines. If you're preparing for RAG interviews, this is a very helpful resource.
    AnonymousRAG CandidateIdentity protected
The team

Written by people who
ship this for a living.

Everything here comes from someone who does the work. Nothing is licensed in, nothing is ghostwritten, and every piece carries the name of whoever wrote it.

Naresh EdagottiFounder · AI engineer
If I cannot show you the line where it breaks, I have not understood it well enough to write about it.
Writes the deep dives and builds the interview kits. Everything here started as his own interview notes.What attention heads actually learnTransformers, Illustrated
Placeholder nameCurriculum · evaluation
A course that never tells you what you got wrong is entertainment. Evaluation is the part that makes it teaching.
Owns what a course teaches and in what order, and the evaluation sets that keep it honest.Evaluation & ObservabilityA golden set you can defend
Placeholder nameSystems · infrastructure
Most latency is queueing, and most cost is a retry nobody logged. Neither shows up in a notebook.
Keeps the serving path fast and the costs legible. Writes the production and cost material.Cost & Latency EngineeringA serving path that holds its P99
Placeholder nameCareer practice
The question is never whether you know it. It is whether you can still explain it while someone watches.
Runs the mock interviews and the resume work, and turns what comes back into new kit questions.The AI Engineer kit1:1 AI Career Guidance
Three of the four records above are placeholder people, and all four position quotes are draft copy written for this mockup - nobody named here has said them. Social links go nowhere. Replace with real colleagues and real words before this goes anywhere public.
Career services

Into an AI role,
with a person in your corner.

Nine one-to-one engagements for students and working professionals. Decide what to aim at, build proof worth showing, then practise being asked about it.

Scoped, then quotedNo fixed prices. Every engagement is quoted per scope.
No enrolmentA conversation, then the work. Not a seat on a course.
How an engagement runs
  1. 01You enquire. Tell us where you are and what you are aiming at.
  2. 02We scope it. The work is agreed, then quoted for that scope.
  3. 03It gets booked. A person's time, not a seat on a course.
Book an intro call
9 engagements · 3 stages · quoted per scope
Stage 01

Direction3 engagements

Know what to aim at before you spend a single evening studying.

  • 1:1 AI Career Guidance60-min live call
    • Skill gap audit
    • Role-fit shortlist
    • Written action plan
    • A follow-up two weeks later
    1 artefact you keep
  • AI Learning RoadmapCustom plan + review call
    • 12-week schedule
    • Curated resources
    • Checkpoint milestones
    • One rescope when life happens
    2 artefacts you keep
  • Career Transition CoachingMonthly engagement
    • Transition plan
    • Positioning story
    • Monthly check-ins
    2 artefacts you keep
Stage 02

Proof3 engagements

The portfolio and profile that make a recruiter stop scrolling.

Stage 03

Reps3 engagements

Practice under pressure, with someone in your corner between rounds.

  • AI Mock Interviews60-min simulated round
    • Live interview
    • Scored feedback
    • Recording to rewatch
    • Follow-up drills
    3 artefacts you keep
  • AI Interview PreparationMulti-session program
    • Topic playbook
    • Question bank
    • Project narrative
    3 artefacts you keep
  • Mentorship Programs3 or 6 months
    • Bi-weekly sessions
    • Async support between them
    • Quarterly goal reviews
    Time with a person, not a deliverable
Not sure which one you need? Start with the intro call. If none of these fit what you are trying to do, we will say so. Book an intro call

Before you start

The five things people ask most.

How often does something new arrive?
A new deep dive every weekday, and a course finishes every few weeks. Nothing goes out that has not been read end to end by someone who works in the field.
Do I need a maths background?
Comfort with linear algebra and Python helps. Where a derivation matters we show it in full rather than pointing at a paper and moving on.
Can I read on my phone?
Yes. The reader is built for long reading on a small screen, and the newsletter carries the full article for members.
Do I keep a course after I have bought it?
Yes. There is no subscription and nothing renews. A course you have bought stays readable, and the parts added to it later come with it at no extra cost.
Does buying one course open the others?
No. Each course is bought on its own. The free concepts, the newsletter and the opening part of every course stay open to everyone either way.
Still unsure? Every concept and the deep dive behind it is free, so you can check before you spend anything. Read today's concept

The newsletter

The same writing,
in your inbox instead.

Every new course, sent the day it is published. Nothing else.