Building with AI: Fundamentals & OpenRouter
✕
Staff Software Engineer with over 9+ years of experience building Android products at scale, Kotlin systems, and AI workflows.
Architecting high-scale Android applications, Kotlin Multiplatform, and resilient mobile systems.
Building with LLM APIs, developer agents, and on-device intelligent systems.
Organizing Kotlin/Android circles in Chennai & sharing deep-dives on YouTube @shravzdev.
A quick puzzle before we dive into how modern AI actually generates text.
Before modern LLMs, Transformers, or chatbots existed... an incredible breakthrough in 1950 changed how we predict information forever.
Before software apps or AI existed, Claude Shannon ran a pen-and-paper statistical experiment at Bell Labs, asking human subjects (including his wife Betty) to guess the next letter in covered sentences.
Shannon mathematically proved that given context like "THINK FAST. BUILD FASTER. SHIP TODA...", the next character is statistically predictable. This theoretical paper founded the principle modern LLMs use today.
How Shannon's 1950 theoretical guessing experiment evolved into modern token & word prediction engines.
Shannon tested human subjects guessing covered letters to prove language is statistically predictable: "SHIP TODA[ Y ]".
Modern LLMs predict sub-word tokens (~4 chars) across a 100k+ vocabulary to stream entire answers.
The Idea: What if a simple switch could imitate a biological brain cell?
Real-World Change: Shifted computers from calculators into machines wired like brains.
The Idea: If you talk to a machine behind a curtain and can't tell it's a computer, is it intelligent?
Real-World Change: Created the world's first benchmark for human-level AI.
The Idea: Claude hid letters in sentences and Betty guessed what came next. She was almost always right!
Real-World Change: Proved language is statistically predictable—the engine of ChatGPT.
The Idea: Co-organized by Claude Shannon, John McCarthy & Marvin Minsky to define machine learning.
Real-World Change: Coined "Artificial Intelligence" and formally founded AI as an academic discipline.
The Idea: Programmers manually wrote strict rules: "IF user says hello, THEN say hi".
Real-World Change: Worked for chess, but broke on messy human language and slang.
The Idea: Parroted words back: "I'm feeling sad" → "Why are you sad?" People poured their hearts out!
Real-World Change: Proved humans love talking to bots, but exposed zero true understanding.
The Idea: Doctors and banks fed 500-page rule books into computers to diagnose illness and approve loans.
Real-World Change: Showed software could assist, but crashed whenever rules conflicted.
The Idea: When rule bots proved too brittle, companies froze funding for a decade.
Real-World Change: Forced scientists to invent Machine Learning: learning from data instead of hand-coded rules!
The Idea: Like a kid shooting basketballs, every time the neural net misses, it tweaks connections until it scores.
Real-World Change: Gave neural networks the superpower to self-correct automatically.
The Idea: Instead of describing "a cat has whiskers", engineers fed the computer 100,000 labeled photos.
Real-World Change: Enabled email spam filters, postal handwriting scanning, and fraud alerts.
The Idea: Researchers trained deep neural networks on fast Nvidia gaming GPUs using millions of photos.
Real-World Change: Ushered in Deep Learning, unlocking Face ID and self-driving cars.
The Idea: Billions of smartphone users generated endless photos, videos, and voices for AI to learn from.
Real-World Change: AI moved out of academic laboratories and directly into every pocket on Earth.
The Idea: Transformers look at entire paragraphs at once, understanding deep context.
Real-World Change: Created Large Language Models, scaling reading speed by 1,000x!
The Idea: Claude Shannon's 1950 letter-guessing game scaled across billions of parameters predicting the next word.
Real-World Change: AI transformed from search engines into creative partners that write code, poetry, and apps.
RLHF: Reinforcement Learning from Human Feedback.
The Idea: Raw internet text is messy. Human teachers graded AI replies to guide it to be polite, honest, and helpful.
Real-World Change: Turned raw probability engines into safe, friendly AI assistants anyone can talk to.
The Idea: You don't need a supercomputer. A single API key connects your app to any model out there.
Real-World Change: Anyone in this room can build and ship a full AI startup in hours!
By the end of this you’ll know what an LLM is, how it answers you, and the handful of terms you’ll use all day — in plain words, no maths.
How AI, ML, Deep Learning & LLMs fit cleanly together.
Tokens, next-token prediction loops, and context windows.
6 concrete architectures for wiring AI into real apps.
The broad idea: machines doing things that normally need human smarts.
Systems that learn patterns from data instead of being hand-coded with rules.
ML using large neural networks — the engine behind modern AI.
Deep-learning models trained on huge text to predict and generate language. This is what you’ll be calling today.
The prompt you send — a question or instruction.
Text is chopped into tokens (word-pieces) the model can read.
The model scores every possible next token and picks one.
It adds that token and predicts again… and again.
The trained AI you call (e.g. GPT-4o, Claude, Llama). Different models = different skill, speed, cost.
A word-piece. Prompts and replies are measured — and billed — in tokens.
The max tokens a model can hold at once. Prompt + reply must fit inside it.
Creativity dial. Low = focused and repeatable, high = varied and surprising.
When a model states something wrong with full confidence. Always sanity-check facts.
Hidden instructions that set the model's role and rules before the user types anything.
Six patterns cover almost every app you could build. Pick one of these instead of just “adding AI”.
Answer questions, act as a helper or tutor inside the app.
Write captions, stories, replies, ideas, quiz questions.
Turn long text or notes into a short, clean summary.
Sort input into categories — mood, topic, spam / not spam.
Pull clean JSON or fields out of messy text for the UI.
Describe or read an image the user captures on the iQOO.
How it plugs in: The app sends text to a model over an API, gets a reply back, and shows it in the UI. That’s the whole loop.
Here you’ll grab an API key with credit and make your first call to a model — before wiring it into the app. One key, many models.
An API in one picture: Your app (customer) & the API (waiter).
Instant team credit from the Reskilll portal with your email.
Let AI tools write the call, pick minimax, and protect credit.
No forms, no waiting. The portal issues your team's OpenRouter key the moment it verifies your hackathon email.
You won’t hand-write API code. An AI coding tool (Claude Code, Cursor, and the like) writes and runs it — you just describe what you want and hand it three things.
https://openrouter.ai/api/v1minimax/minimax-m3 (or openrouter/auto)
Your key unlocks 200+ models. You don’t need to be an expert — pick a sensible default and let AI do the choosing.
openrouter/auto to auto-route each request, or just ask the AI which model fits your feature.
Here you’ll turn a vague idea into a precise prompt, then wire it into an Android app that builds. This is where the app comes alive.
Role, context, task & JSON format that the app can parse directly.
Open starter kit, let Gradle sync, add INTERNET permission & hit Run.
Enable USB debugging, install the debug APK & demo live on hardware.
AndroidManifest.xml early.
An iQOO device won’t accept a hand-built APK until developer mode is on. Do this once, early — before the build is even ready.
Settings → About phone → tap “Build number” 7 times. You’ll see “You are now a developer”.
Back in Settings → System → Developer options → enable USB debugging.
In the same menu, allow “Install via USB” / apps from USB so the APK isn’t blocked.
Plug the iQOO into the laptop. On the iQOO, tap “Allow” on the debugging prompt.
Cleanest way — installs straight from the laptop.
app-debug.apk to the iQOO (USB / Drive / share)
Most blockers are one of these six. Find the symptom, apply the fix, and keep building.
Key is wrong or missing. Check the Authorization: Bearer header and that the key was copied whole.
Add the INTERNET permission to AndroidManifest.xml, then rebuild.
Usually network. Retry sync, check Wi-Fi, and make sure the SDK finished downloading.
Re-plug the cable and tap Allow on the iQOO’s debugging prompt. Toggle USB debugging off/on.
Uninstall any old copy first; enable install-from-USB / unknown apps on the iQOO.
Check the OpenRouter dashboard — switch to a cheaper model or slow the calls down.
AI → A working Android app, shipped to an iQOO device.
Clone starter repo template & configure your project setup.
Claude Code, Cursor & Gemini write and iterate your feature code.
One door to 200+ models with your team's $10 event credit.
Build native Kotlin UI with Jetpack Compose & wire live API data.
Deploy APK to physical iQOO devices & present your live demo!
Scan the QR code to explore tutorials, open source projects, and dev content.
