Career talk

From ECE
to GenAI

Becomingp 0.41 ap 0.93 softwarep 0.38 engineerp 0.72 inp 0.55 thep 0.81 agep 0.34 ofp 0.97 AIp 0.46 agentsp 0.29

Shishir Subedi · Sr. AI and Backend Engineer @ QsystemsAI

→ / Space: next · ←: back · N: speaker notes · T: theme · F: fullscreen

My journey

I didn't start in computer science

Electronics & Communication Engineering
Where I started. Circuits and signals, not code.
Data Mining elective
The course that changed my direction.
More CS electives
AI, Databases and more in 3rd and 4th year.
Final project in Python
My first real thing built end to end.
Python intern
Learning how real software teams work.
Full Stack Python Developer
7+ years of Python since then.
Master's in CS: Data Science & Analytics
Had to stop in the final semester. The learning stayed with me.
2023: ChatGPT API arrives
Built my first GenAI project with no AI framework.
GenAI for a US bank
Through a Nepal company. A private ChatGPT-like platform.
Research paper published
Tabular Embedding Model (TEM).

What I learned from it

I didn't have a perfect plan. I followed what made me curious, and I kept building.

Curiosity→ Elective→ Project→ Internship→ Job→ Research

Not everything went to plan. I had to leave my master's in the final semester. My career kept growing anyway, because I kept building.

If you're in 1st year and still confused about what you want, that's normal. Pick something, build it, and see what you enjoy.

Part 1 · What is an LLM?

ChatGPT is a very good next-word predictor

It reads the words so far, scores every possible next word, picks one, adds it, and repeats. It does this so well that we often can't tell AI writing from human writing.

Words so far"The capital of Nepal is"

The modelScores every word piece it knows

Probabilities

Kathmandu

a

located

Pick one"Kathmandu"

↺ Add the word to the text, then predict again, until the answer is finished

It generates. It does not check. That's why it can sound confident and still be wrong.

Game 1 · You are the model

Round 1 of 6

Guess the next word

The capital of Nepal is ?

Shout your guess first. Then reveal what a model would predict.

Numbers are illustrative, shaped like a real model's output.

Game 2 · Temperature

Turn up the temperature, and rare words get a chance

?

Temperature 1.0

Bar: chance at this temperature. Dark tick: the model's normal chance (temperature 1). Right column: how many times each word was picked. Numbers are illustrative.

Game 3 · A tiny language model, running live

Watch a model write, one word at a time

Temperature 1.0

1.0 picks words in proportion to how often it saw them.

Candidates for the next word (looking at the last 2 words: )

What it learned from: sentences

Same idea, much bigger

Our toy model vs a real LLM

	Our toy model	A real LLM (ChatGPT, Claude, Gemini)
Learned from	23 sentences	Trillions of words: books, websites, code
Looks at	The last 2 words	The whole conversation and long documents
Knows	75 words	Many languages, code, math, Nepali too
Learns by	Counting	A neural network with billions of numbers, tuned with calculus
Job	Predict the next word	Predict the next word

Early ChatGPT (2023) could only do this: answer from what it learned in training. No internet, no actions. Then things changed.

Part 2 · AI agents

Generative AI talks. An AI agent acts.

Generative AI (ChatGPT, early 2023)

You: "How do I buy a gaming mouse?"

LLM

Text: "Go to a store website, search…"

Only gives answers. Knows only what it was trained on. You do the work.

AI agent (LLM + tools + a loop)

Web search

Browser

Run code

Files

LLMthe brain

Email

Database

Calendar

Payments

You: "Buy me a gaming mouse under Rs. 5,000." It searches, compares, adds to cart and pays for you.

Brain = the LLM. Hands = the tools. The loop = it keeps working until the job is done.

How an agent works

An agent is a `while` loop with a brain

while (!task_done) {    thought = think(goal, history);    result  = act(thought);   // use a tool    history = observe(result);}report_to_user();

Goal: "Buy me a gaming mouse under Rs. 5,000."

Safety

Give agents a key, not the master key

Agents act on your behalf, and they can make mistakes. Give them only the access the task needs.

Least access. Read-only wherever possible.
A human approves anything that spends money or deletes data.
Never give an agent production access. No live databases, no real credentials.

Shopping agent permissions

Search the webAllowed

Browse the storeAllowed

Add to cartAllowed

Pay with saved cardAsk human first

Change account passwordNever

Company production databaseNever

Part 3 · A real project

GenAI for a US bank, built from Nepal

1

Research first

The bank did not adopt AI at first. We tested how it performs on banking data and whether it was safe to use.

2

Then build

A private ChatGPT-like app for the bank: web search, data collection from third-party sites for analysis, and deep research.

3

Deploy at scale

Full stack, deployed on scalable cloud infrastructure so many employees could use it.

In serious companies, "can we trust it?" comes before "can we build it?"

What is RAG? · Part 1 of 2

RAG: an open-book exam for the LLM

An LLM has never seen your college handbook. RAG (Retrieval-Augmented Generation) finds the right pages and adds them to the prompt.

Indexing pipeline · done once RAG pipeline · every question (next slide)

Example handbook, made up for this talk.

What is RAG? · Part 2 of 2

The RAG pipeline: answering a question

Indexing pipeline · done once RAG pipeline · every question

The LLM never learned your handbook. It read the right page at the moment you asked.

The problem we hit

RAG worked for text. Tables were a problem.

Chunking works when each piece makes sense on its own. A paragraph does. A table row doesn't.

Text documents

Works

PDF→Small chunks→Find relevant→LLM answers

chunk · page 7"Students must have at least 80% attendance to sit for the board exam."

One chunk holds a complete idea. The LLM can answer from it directly.

chunk · rows 100231–100233100231, Branch 12, 38,500 · 100232, Branch 03, 12,000 · 100233, Branch 07, 55,250

A chunk of table rows is a few random numbers. What's the average loan in 2024? No single chunk can tell you.

Tables (CSV, database)

Breaks

loan_id	branch	amount	issue_date
100231	Branch 12	38,500	2024-02-11
100232	Branch 03	12,000	2024-02-11
100233	Branch 07	55,250	2024-02-12
… 1,199,997 more rows

Too big to paste into the prompt. Cutting rows into chunks breaks totals and averages.

Example table. Not real bank data.

Our research · Tabular Embedding Model (TEM)

Index the menu, not the kitchen

Analogy

A bank manager doesn't read a million transactions. They ask an analyst to run the numbers, then decide from the summary.

Rows the LLM reads

Example question and numbers, for illustration.

Part 4 · How I work today

I haven't seriously typed code in the last 1.5 years.

Why would I? AI writes it faster than me.

But I read, review and question every line. I can do that only because I wrote code by hand for years before.

My rules for working with AI

AI writes the code. I stay responsible for it.

1
Always reviewIt makes mistakes and adds things nobody asked for. It's a generator, so it keeps generating.
2
Know the processTesting, security, code review, deployment. AI doesn't replace the process.
3
No production accessNever give agents live databases or real credentials.
4
Make it explain itself"Why did you write this? Why not the other way?" If I don't understand it, I don't ship it.
5
Document the whySo the next agent, or human, fixing this code knows why it was built this way.

Part 5 · The landscape has changed

The lines between roles are disappearing

Frontend, backend, DevOps and AI used to be separate jobs. With AI doing much of the typing, companies want engineers who can own the whole problem.

Talking to the client

Backend

DevOps & Cloud

Security

AI & Data

Frontend

FDEForward Deployed Engineer

A role you'll hear about more

Forward Deployed Engineer

An engineer who works directly with the customer, understands their real problem, and builds the whole solution end to end. The role started at Palantir. Today AI companies like OpenAI and Anthropic hire FDEs.

Example: my bank project was FDE work.

Understand

What does the bank actually need?

Research

Is AI safe and accurate on their data?

Build

Backend, frontend, AI pipeline.

Deploy

Scalable cloud, in production.

Judged by one thing: did the customer's problem get solved?

BIM students: this role suits you. Understanding the business problem is half the job.

Part 6 · How to start

Use AI as much as you can, as a tutor

Copy machineDone tonight, learned nothing

TutorBetter at the next one

How?

Write a C program to reverse a string.

How do I reverse a string? Give me a hint, not the full code.

Why?

Fix this error.

Why did I get this error? Explain it like I'm new to C.

Why not?

Write my factorial program.

You used a for loop. Why not a while loop?

Same homework on both sides. Keep asking how, why and why not. Interviews don't allow ChatGPT.

The foundation

Build from the bottom up

Mathlinear algebra · probability & statistics · calculus
C → C++memory, pointers, OOP. Then any language is easy.
DSA + DBMSjudge if code is efficient · every app stores data
Git · Linux · Cloudhow real software is shared, run and deployed
Python · JavaScriptbuild and ship
AI apps & agents

AI builds fast. Foundations keep it standing.

A house without a foundation doesn't survive the first earthquake. Neither does your code.

If you want to do serious AI work

Your math classes are inside ChatGPT

Linear algebra

Words become vectors (embeddings). Similar meaning = nearby points. TEM uses this.

Probability & statistics

Kathmandu

a

located

also

Next-word prediction is probability, learned from the statistics of huge amounts of text. Temperature reshapes the distribution, as you saw in Game 2.

Calculus

Models learn by gradient descent: step downhill along the slope. It can get stuck in a local minimum.

You already have these in your syllabus. They are what's inside ChatGPT.

You don't need math to use AI. You need it to build AI or go into research.
The ball is gradient descent: each step moves opposite to the slope. On entering the slide it starts on the right and gets stuck in the local minimum.
Analogy: walking down a hill in thick fog. You only see your feet. You reach a small valley and think you're at the bottom.
Click "Start left": it reaches the global minimum. Same method, different starting point, very different result.
Real training uses tricks to avoid getting stuck: momentum, random starting weights, and noise from mini-batches.
How to read the curve: left to right = different settings of the model's weights (here 1 weight, in ChatGPT billions). Height = error, how wrong the model is. Training = finding the lowest point.
Global minimum = the best possible settings, lowest error anywhere. Local minimum = settings where every small change makes things worse, so the model stops, but a much better setting exists somewhere else.
Example, spam filter: stuck in a local minimum, it learns "emails with the word FREE are spam". It catches some spam and stops improving, because small tweaks don't help. At a better minimum it learns many signals together: strange links, unknown sender, urgent tone.
Example, Pathao or Google Maps: at every junction you take the road that looks shortest right now, and end up stuck in Kalanki traffic. Every local choice looked good, but the best route overall was a different road. Choosing the best-looking small step every time can trap you.
Example, career (ties back to your story): a comfortable path where every small change looks worse is a local minimum. Sometimes you need a big jump to reach a better valley, like I did moving from ECE to computer science.
How the fixes map: momentum = a ball rolling with speed carries over small bumps. Random starts = try from many places and keep the best. Learning rate = step size: too big and it jumps around, too small and it takes forever.
Honest note if someone asks: in very large networks, most local minima turn out almost as good as the global one. The bigger practical problems are flat areas where the slope is nearly zero and learning slows down. Nobody finds the true global minimum for ChatGPT. The goal is an error low enough to work well.

Calculus in action · the base of deep learning

Backpropagation: how a neural network learns

Guess, measure the error, use derivatives to see which way each weight should move, then adjust. Repeat millions of times.

Ready

Round	w	Prediction ŷ	Error L
No rounds yet

Task: learn the price of one plate of momo. 2 plates cost Rs. 300. The model starts with a bad guess, w = 100.

A negative derivative means "increase w to reduce the error". After a few rounds w reaches about 150, the real price per plate. Nobody told it 150. It learned from data.

forward →input→layer 1→layer 2→… layer 96→error

← backward: chain rule through every layer, for billions of weights

PyTorch and TensorFlow calculate these derivatives automatically (autograd). Knowing what they do is what lets you debug training.

Click Forward pass: the model predicts 2 × 100 = 200. Real answer 300. Error = 10,000.
Click Backward pass: this is the chain rule from your calculus class. dL/dw = dL/dŷ × dŷ/dw. The red arrows show the error flowing back to the weight.
Click Update weight: w moves opposite to the derivative, 100 → 120. Learning rate 0.05 controls the step size.
Then click Train 10 rounds and watch the error column drop to nearly zero while w approaches 150.
Big picture: ChatGPT trains the same way, with the same chain rule applied backwards through many layers and billions of weights. This is why calculus matters for deep learning.
Why square the error? 1. Signs don't cancel. Predicting 200 for a Rs. 300 order is off by −100. Predicting 400 is off by +100. Both are equally wrong. Without squaring, averaging −100 and +100 gives 0, which looks perfect. Squared, both are 10,000.
2. Big mistakes cost much more. Off by 10 → error 100. Off by 100 → error 10,000. The model works hardest on fixing its worst mistakes.
3. Easy, smooth derivative. The derivative of (ŷ − y)² is 2(ŷ − y): the slope is proportional to how wrong you are. Far from the answer, big steps. Close to it, small careful steps, automatically. That's why w moved 100 → 120 → 132 → … and slowed down near 150.
Why not just |ŷ − y| (absolute error)? It has a sharp corner at zero, where the derivative jumps from −1 to +1. The slope stays the same size however close you are, so it tends to overshoot. Squared error is a smooth bowl.
Statistics link: minimising squared error finds the average (mean). It's the same idea as least squares in your statistics course.
If someone asks "Does ChatGPT use squared error?" No. Squared error is for predicting numbers: prices, temperatures, marks. For choosing the next word, models use cross-entropy loss, which punishes giving a low probability to the correct word. The backprop and chain rule part is exactly the same.
Downside to mention if asked: because big errors are squared, one crazy data point (an outlier, like a typo of Rs. 30,000 instead of 300) can pull the whole model towards it.

What a deep neural network looks like

Layers of neurons, connected by weights

Forward: words go in, a prediction comes out

Every line is a weight the model learns. This drawing has about 600. Models like ChatGPT have billions, across many more layers.

Your roadmap

What to do, year by year

1st Build the base

C, then C++. Really understand them.
Git and GitHub from day one
Basic Linux commands
Ask AI "why", not "write"

2nd Start building

DSA and DBMS, seriously
2–3 real projects: college notice app, hostel management, a shop's inventory
Deploy them. A live link, not a zip file.

3rd Go deeper

Pick a direction: web, mobile, backend, data, AI
One project using an AI API, like a chatbot over your syllabus PDFs
Cloud basics, internships, open source

BIM

Your edge is business + tech. Business analyst, product roles, AI automation for companies, data analysis, and FDE-style roles all need people who understand both the problem and the tool.

Practical tools

Learn the tools real teams use

# Git: save and share your work
$ git init
$ git add .
$ git commit -m "first project"
$ git push origin main

# Linux: servers run Linux
$ ssh user@my-server
$ ls   cd   cat   grep   chmod
$ tail -f app.log   # watch live logs

Cloud: AWS, Google Cloud, Azure

New accounts get free credits or a free tier. Run a small server and a database, and deploy your project for free for a while.

GitHub Student Developer Pack

Free tools and cloud credits for students. Sign up with your college email or student ID.

Set a billing alert on day one so a forgotten server doesn't surprise you.

What you'll deploy one day

A typical e-commerce app on AWS

Click "Follow a request" and walk through the 6 steps: browser, DNS, load balancer, servers, database, images.
Analogy: Route 53 is the phone book, the load balancer is the receptionist sending you to a free counter, the servers are the staff, the database is the locked record room, S3 is the warehouse for photos.
Security point: the database sits in a private subnet with no internet access. Only the backend can talk to it. That's the same "least access" idea as the agent safety slide.
Scaling point: during a Dashain sale, add more servers behind the load balancer. Users never notice.
If someone asks: in production many teams also put the frontend and backend servers in private subnets and keep only the load balancer public, and put CloudFront (a CDN) in front of S3 for faster images.
Link to FDE: one engineer who understands this whole picture, not just one box, is very valuable.

If you remember only four things

1
AI writes the code. You stay responsible for it.
2
Fundamentals (C, DSA, DBMS, math) let you judge what AI gives you.
3
Use AI as a tutor. Ask how, why and why not.
4
Build, deploy and show your work on GitHub.

Questions?

Thank you · Shishir Subedi

From ECEto GenAI

I didn't start in computer science

ChatGPT is a very good next-word predictor

Guess the next word

Turn up the temperature, and rare words get a chance

Watch a model write, one word at a time

Our toy model vs a real LLM

Generative AI talks. An AI agent acts.

Generative AI (ChatGPT, early 2023)

AI agent (LLM + tools + a loop)

An agent is a while loop with a brain

Give agents a key, not the master key

Shopping agent permissions

GenAI for a US bank, built from Nepal

Research first

Then build

Deploy at scale

RAG: an open-book exam for the LLM

The RAG pipeline: answering a question

RAG worked for text. Tables were a problem.

Text documents

Tables (CSV, database)

Index the menu, not the kitchen

AI writes the code. I stay responsible for it.

The lines between roles are disappearing

Forward Deployed Engineer

Use AI as much as you can, as a tutor

Build from the bottom up

Your math classes are inside ChatGPT

Linear algebra

Probability & statistics

Calculus

Backpropagation: how a neural network learns

Layers of neurons, connected by weights

What to do, year by year

1st Build the base

2nd Start building

3rd Go deeper

Learn the tools real teams use

Cloud: AWS, Google Cloud, Azure

GitHub Student Developer Pack

A typical e-commerce app on AWS

Questions?

Speaker notes

From ECE
to GenAI

An agent is a `while` loop with a brain