RAJIV GANDHI UNIVERSITY OF KNOWLEDGE TECHNOLOGIES · BASAR
From Idea to
Deployed AI
A hands-on course in AI engineering. You will build one
application end to end and put it on the internet, add presence
verification and face matching, give it an AI agent over your own
data, and measure whether it actually works.
About This Course
Two objectives, and how the sessions run.
OBJECTIVE ONE OBJECTIVE TWO
Know what is actually happening in this Finish holding something real.
industry right now. A live URL a stranger can open. A repository with your
AI engineer, agent, harness, context engineering, evals, name on it. A user who is not your classmate.
forward deployed engineer, spec-driven development — Interviews for these roles do not test whether you
this vocabulary is eighteen months old or less, and it is memorised a framework — they ask what you built, why
what job descriptions are written in this year. Most you built it that way, and how you knew it worked. You
graduates have heard the words. Very few can say what will have honest answers to all three.
they mean. You will have done all of them.
We are not training models. We are doing AI engineering: building products on top of models that already exist. The hard parts
are not mathematics — they are problem framing, deployment, tool design, and knowing whether the thing actually works.
How each session works: the 3H method
BLOCK TIME WHAT HAPPENS
Head ~12 min The concept. Whiteboard, diagrams, and the smallest amount of theory that makes the practical work
make sense.
Heart ~6 min Why it matters. What the industry is actually doing, what the job looks like, and where the hype ends.
Hands ~40 min You build. Every session ends with something working on your own machine or your own server.
The Hands block is the point of the course. If a concept runs long, we cut the concept — not the building. The class hour is followed by a
required lab hour, which is where the learning consolidates.
One application, eleven layers
You will not build eleven unrelated demos. From Day 2 you build one application — a student attendance system this class
genuinely needs — and it survives to the end. The app is the battlefield: it is where each new practice gets tested. The
practices are the point.
What you will walk away with
A live public URL anyone can open, over HTTPS An LLM feature whose cost per call you can state in
A GitHub repository with a README you wrote rupees
An app that survives a reboot and keeps its secrets An agent loop you wrote by hand and can draw from
out of git memory
A presence check you can explain and attack An evaluation set from a real user's questions, and a
score that moved
A face-match threshold you chose from measured
data A named stakeholder who has actually used your
software
A one-page write-up that is your resume bullet, blog
post and demo script at once
Reading for every session is listed on that day's entry. Required items are capped at roughly 15 minutes per day; optional items are
for students who want to go further.
Course Structure
What the app gains — and what you will be able to name afterwards.
DAY THEME THE APP GAINS THE PRACTICE YOU LEARN TO NAME
1 Programming, — Decomposition. Using an LLM as a tutor, not a vending machine
Recapped
2 Say It Before You Skeleton runs, on GitHub Spec-driven development · agentic coding
Build It
3 Where Data Comes Real students in the Context engineering · scraping law and ethics
From database
4 Ship It to the Internet Live HTTPS URL Reverse proxies · secure contexts · TLS
5 Make It Survive Survives reboot, secrets Harness engineering · "you build it, you run it"
safe
6 The 75% Question Query functions + a real Forward Deployed Engineering · the customer interview
user
7 Proof You Were Layered presence checks Spec-anchored development · never trust the client
There
8 Proof It Was You Face matching + consent Embeddings · thresholds as product decisions
controls
9 Put an LLM In It Plain English → a real Tokens · context window · structured output · cost in ₹
answer
10 The Agent Loop An admin agent over your Agents · tool design · MCP · Skills · loop engineering
data
11 Evals & Attacks An eval set and a score Evals · LLM-as-judge · prompt injection · lethal trifecta
that moved
From Day 6, your project stops being everyone else's
On Day 6 each pair interviews a real person on this campus — a warden, a lab in-charge, the placement cell — and
builds for their questions. From Day 9 your agent has different tools from your neighbour's, and on Day 11 your evaluation
set is their twenty real questions. Thirty pairs, thirty genuinely different repositories — which is what makes this a portfolio
piece rather than a class exercise.
Before Day 1
Work through this the day before — not the morning of. Every item takes minutes when it works and an hour when it does not.
DONE ITEM HOW YOU KNOW IT WORKED
☐ Python 3.11 or newer python3 --version prints a version number
☐ git git --version prints a version number
☐ A code editor VS Code, or whatever you already like
☐ GitHub account You can log in and create an empty repository
☐ Terminal comfort You can cd into a folder and list its contents
☐ Your phone on campus WiFi You can open a web page on it
Prerequisites: you can write and run a Python script, and you have used git at least once, even badly. No machine learning,
mathematics or cloud experience is required.
The Eleven Days
What each session builds, the vocabulary it gives you, and what to read beforehand.
D AY 1 Programming, Recapped
You build: a map of your own gaps, a decomposition of the app you are about to spend ten days building, and a working
habit for closing gaps on your own.
decomposition state abstraction interface invariant Software 1.0 / 2.0 / 3.0 comprehension debt
SETUP Install Python 3.11+, git and VS Code; create a GitHub account. ~30 min — the day before
READ The Batch — subscribe, read the most recent opening letter. [Link]/the-batch · 5 min
WATCH Andrej Karpathy — "How I use LLMs". The best demonstration of using these tools well. YouTube · 2h 11m
READ Ethan Mollick — One Useful Thing. The clearest writer anywhere on working with AI. [Link] · 15 min
BROWSE MIT — The Missing Semester of Your CS Education. Shell, git, debugging. [Link]
D AY 2 Say It Before You Build It
You build: a working local web app in your own GitHub repository, and the first specification you have ever written.
spec-driven development spec-first spec-anchored spec-as-source acceptance criteria EARS notation agentic coding
Spec Kit Kiro
READ GitHub — "Spec-driven development with AI". The workflow, from the team that built Spec Kit. [Link] · 12 min
READ Birgitta Böckeler — "Understanding Spec-Driven Development: Kiro, spec-kit, and Tessl". Read this one. The best
thing written on the subject, and notably unconvinced. [Link] · 20 min
READ Anthropic — "Claude Code: Best practices for agentic coding". How the people who build these tools actually use
them. [Link]/engineering · 20 min
BROWSE Amazon Kiro — spec docs. [Link] → [Link] → [Link] , and EARS notation. [Link]/docs/specs ·
10 min
D AY 3 Where Data Comes From
You build: a polite scraper, and a database with real students in it.
DOM CSS selector [Link] rate limiting CFAA hiQ v. LinkedIn context engineering just-in-time retrieval
READ hiQ Labs v. LinkedIn. Read to the end — the second ruling matters as much as the first. Wikipedia · 10 min
READ Anthropic — "Effective context engineering for AI agents". Compaction, note-taking, sub-agents, just-in-time retrieval.
[Link]/engineering · 15 min
READ Jason Liu — "Grep Beats Embeddings". Why the simple retrieval method beat the clever one. [Link] · 10 min
READ The [Link] of a site you use every day. Notice what it asks, and who enforces it. 5 min
D AY 4 Ship It to the Internet
You build: your application, live on the public internet over HTTPS, at a URL you can send to your family — with a working
camera.
port process reverse proxy DNS secure context TLS certificate authority Let's Encrypt
READ MDN — "Secure contexts". Read the list of restricted features; you are about to be bitten by two of them.
[Link] · 5 min
SETUP Confirm you can ssh to the class server with the credentials issued on Day 3. 10 min — the commonest reason students
fall behind today
READ "What is a reverse proxy?" — the Cloudflare Learning Center explainer. [Link]/learning · 5 min
READ MDN — [Link] . In an insecure context the property is absent, not denied. [Link] · 5
min
D AY 5 Make It Survive
You build: an app that survives a reboot, keeps its secrets out of version control, and that you can debug from its logs.
process manager systemd environment variable twelve-factor observability harness engineering you build it, you run it
READ The Twelve-Factor App — Factor III: Config. One short page; the canonical statement of why config lives in the
environment. [Link]/config · 5 min
CHECK Your Day 4 URL is still live. If not, write down what you think happened before you look. 2 min
READ Anthropic — "Harness design for long-running application development". Today's new word, from the people who
coined its current usage. [Link]/engineering · 15 min
READ Anthropic — "Building a C compiler with a team of parallel Claudes". What a serious harness looks like pushed hard.
[Link]/engineering · 15 min
READ Birgitta Böckeler — "Harness Engineering: first thoughts". The sceptical read on the same idea. [Link] · 15
min
D AY 6 The 75% Question
You build: four query functions that bound everything your AI will ever be able to answer — and, in the lab, a real stakeholder
with ten real questions.
groupby aggregation Forward Deployed Engineer product engineer discovery one customer, many capabilities
READ Gergely Orosz — "What are Forward Deployed Engineers, and why are they so in demand?" The single most relevant
career reading of this course. Read it before your interview, not after. [Link] · 15 min
READ PostHog — "WTF is a forward deployed engineer?" Shorter, more concrete, healthily sceptical about title inflation.
[Link]/blog · 10 min
READ Latent Space — "Forward Deployed Engineers and the future of software engineering." Why product and forward
deployed engineering are converging. [Link] · 12 min
READ Eugene Yan — "Patterns for Building LLM-based Systems & Products". Long, canonical, the best map of everything
you build in Days 9–11. [Link] · 66 min
D AY 7 Proof You Were There
You build: a presence check that layers two forgery-resistant signals and honestly records a third it cannot trust.
never trust the client server-side validation X-Forwarded-For TTL replay attack audit log threat model
spec-anchored
READ MDN — Geolocation API, [Link] . It is a 95% confidence radius in metres, and the specification promises
nothing about how it was obtained. [Link] · 5 min
READ Böckeler on spec-driven development — if you skipped it on Day 2, read it today. It lands harder once you have had to
rewrite a specification because physics disagreed with it. [Link] · 20 min
READ "Structured-Prompt-Driven Development". "When reality diverges, fix the prompt first — then update the code."
[Link] · 20 min
D AY 8 Proof It Was You
You build: face verification with a threshold you can defend, an attack you have measured, and consent controls that work.
embedding vector cosine similarity threshold false accept / false reject liveness biometric DPDP Act consent
READ India's Digital Personal Data Protection Act — what counts as personal data and what consent requires. You are
processing biometric data belonging to identifiable classmates. any reputable summary · 10 min
READ Vicki Boykis — What Are Embeddings. A free book, and the clearest explanation of today's central idea anywhere.
[Link]/what_are_embeddings · free PDF
READ Anthropic's Acceptable Use Policy. Note the prohibition on facial recognition — it is why today's matching runs on a
local model, not an API. [Link]/legal/aup · 10 min
SKIM ArcFace (arXiv 1801.07698) — the paper behind the model you use today. Skip the mathematics; read the abstract, look
at the figures. [Link] · 10 min
D AY 9 Put an LLM In It
You build: a live AI feature on your public URL, whose cost per call you can state out loud.
token context window statelessness temperature structured output JSON schema cost per call
prompt → RAG → fine-tune
READ Anthropic — Messages API overview / quickstart. Read the request and response shape; memorise nothing. Find
usage . [Link] · 10 min
READ Eugene Yan — "Prompting Fundamentals and How to Apply them Effectively". Short, practical, better than most
prompt-engineering courses. [Link] · 17 min
BROWSE Claude Cookbooks — runnable notebooks for tool use, JSON mode, evals and prompt caching.
[Link]/anthropics/claude-cookbooks
BROWSE Any provider's pricing page. Note how much cheaper the small models are. 5 min
D AY 1 0 The Agent Loop
You build: a working agent, written from scratch in an empty file, with tools that touch your own application's data.
agent tool tool schema stopping condition workflow vs agent MCP Agent Skills sub-agent
context window as RAM loop engineering
READ Anthropic — "Building Effective AI Agents". The most important reading of the course. Focus on the distinction: a
workflow orchestrates through predefined code paths; an agent directs its own. [Link]/engineering · 15 min
READ Simon Willison — "AI agents can mean a lot of different things". Cuts through the definitional mush in ten minutes.
[Link] · 10 min
READ Anthropic — "Writing effective tools for agents — with agents". Why tool descriptions are prompts.
[Link]/engineering · 12 min
READ Philipp Schmid — "8 Tips for Writing Agent Skills" and "How to correctly use MCP servers". [Link] · 20 min
WATCH Dex Horthy — "12-Factor Agents". Also read Cognition — "Don't Build Multi-Agents", the antidote to a lot of noise.
YouTube · [Link]/blog
D AY 1 1 Evals, Attacks & What Comes Next
You build: an evaluation set, a measured change, and a two-minute demo of everything you shipped.
eval eval set LLM-as-judge golden dataset error analysis prompt injection lethal trifecta exfiltration guardrail
READ Simon Willison — "The lethal trifecta for AI agents". Short, vivid, and the security idea most working engineers still do
not know. [Link] · 8 min
PREPARE Your demo. Two minutes, live URL, no slides. Decide who speaks. 10 min
READ Anthropic — "Demystifying evals for AI agents". If you read one optional item in this entire course, read this one.
[Link]/engineering · 18 min
READ Hamel Husain — "Your AI Product Needs Evals" and "A Field Guide to Rapidly Improving AI Products". "Error
analysis is the single most valuable activity in AI development." [Link] · 45 min
READ Eugene Yan — "Evaluating the Effectiveness of LLM-Evaluators". The definitive treatment of LLM-as-judge.
[Link] · 49 min
How you will be assessed
No written exam. Marks are weighted toward things that exist in public, and things an AI cannot produce for you.
MARKS COMPONENT WHAT IT MEASURES
30 Ships Live HTTPS URL, up at demo time, survives a reboot, no secrets in git
25 Works The eleven daily checkpoints, verified by a script you run yourself
15 Measured An eval set, a before score, an after score, and an honest explanation
15 Explains A README in your own words, plus a live two-question viva on your own code
10 Customer Evidence your named stakeholder actually used it
5 Public One published build-log or blog post linking your repository
The Ten People to Follow
Individual links rot. These people do not. Follow them and you stay current without trying.
This field produces more noise than any other part of software. The most useful thing you can do after this course is pick a
small number of genuinely reliable sources and ignore everything else. Here are ten. Between them they will tell you everything
that matters, usually before it reaches LinkedIn.
WHO WHERE WHY THEY EARN YOUR TIME
Simon Willison [Link] The most reliable practical commentary on what actually works,
updated almost daily. If you follow one person, this is the one. His
tag pages on ai-agents and prompt-injection are running
textbooks.
Anthropic Engineering [Link]/engineering Deep, honest write-ups on agents, tools, harnesses, context and
evals from a team shipping them at scale. Roughly 25 posts; almost
all worth reading.
Andrew Ng [Link]/the-batch Weekly letter. Strategy and perspective for people who build, not
people who speculate. Free.
Gergely Orosz [Link] What engineering jobs actually look like, with real reporting and real
numbers. The best source on the market you are about to enter.
Hamel Husain [Link] The person who convinced the industry that evals matter more than
prompts. Read him before you claim your AI feature works.
Eugene Yan [Link] Long, careful, canonical write-ups on LLM patterns and evaluation.
His "Patterns for Building LLM-based Systems & Products" is a map
of the whole field.
Philipp Schmid [Link] The most practical writing anywhere on Agent Skills, MCP and sub-
agent patterns. Short posts, high signal.
Birgitta Böckeler [Link]/articles/ Consistently, usefully sceptical about AI-assisted engineering. Read
[Link] her whenever you catch yourself getting excited — she is the
correction.
Andrej Karpathy [Link] · YouTube Mental models: Software 1.0/2.0/3.0, the autonomy slider, LLMs as
an operating system. And the best free path into how the models
actually work.
Latent Space [Link] · @aiDotEngineer Where the vocabulary of this field gets coined, and the conference
behind it. Talks are posted free on YouTube.
Also worth your attention
Ethan Mollick ([Link]) — the clearest writer alive on how to actually use these tools.
Vicki Boykis ([Link]) — her free What Are Embeddings is the best explainer of that topic anywhere.
Chip Huyen ([Link]) — AI Engineering is the closest thing this field has to a textbook.
Armin Ronacher ([Link]) — created Flask; credible scepticism about agent design.
Jason Liu ([Link]) — hard-nosed practical writing on retrieval and evals.
Cognition and PostHog — two company blogs worth reading rather than skimming.
Follow ten. Ignore the rest. Being the person on a team who has actually read the primary sources is a bigger advantage than it sounds.
After the Course
Free, hands-on, and each one produces something you can put on a CV. In roughly this order.
WHAT COST / LENGTH WHY IT IS WORTH IT
Publish your build log Free · one evening Do this first. One post on what you built, what broke, what you
measured. Without a famous college on your CV this is the cheapest
credibility there is.
Hugging Face Free · 3–4 hrs/week Four units, hands-on in pre-configured environments so nothing to install,
Agents Course · free certificate a final assignment with a leaderboard, and a certificate you can link. The
best structured next step.
Claude Cookbooks Free · self-paced Runnable notebooks: tool use, sub-agents, automated evals, JSON
anthropics/claude- mode, prompt caching. The fastest way to widen what you can build. If
cookbooks Day 9 felt shaky, start with anthropics/courses → API
Fundamentals.
Write a [Link] Free · an afternoon Agent Skills is now an open standard. Write one for your own project — a
[Link] small, current, demonstrable thing almost no graduate has done.
[Link] — Free · ~10 hrs Reflection, tool use, planning, multi-agent — built from first principles
Agentic AI (Andrew Ng) · 8 graded labs before frameworks. The natural deepening of Day 10.
Berkeley — Free · certificate tiers Lectures from people at OpenAI, DeepMind, Meta and Sierra. Heavier and
Agentic AI MOOC research-flavoured; take it if Day 10 left you wanting more.
Three claims you will meet, and should not repeat
Learn the pattern: a precise-sounding number, widely repeated, every citation eventually pointing at one unsourced
blog post.
"FDE postings are up 800%." Repeated everywhere; no primary dataset exists. Cite the reporting, not the number.
"Loop Engineer" as a job title. A real practice; not an established role. Do not claim it on a CV.
FDE salary figures. From one vendor's marketing blog. Pay here is real and good; those numbers are not evidence.
What we deliberately did not cover
RAG — and note why: your Day 6 SQL queries are the retrieval. Then, in order: building an MCP server, prompt
engineering technique, frameworks like LangChain (avoided deliberately — they hide the loop you needed to see),
multi-agent systems, how the models actually work, and fine-tuning.
One last thing
The most valuable habit in this field is not a technology. It is building something small and writing about it in public,
over and over. Nearly everyone named in this outline built their reputation that way, and none needed permission to start.
Links change often in this field; when something has moved, search the title.