DEV Community: Durva Shah

Thali Edition, Part 3: The Secret Settings of the Kitchen ✅🍱

Durva Shah — Tue, 02 Jun 2026 06:45:00 +0000

Every artificial intelligence system in production is doing the exact same thing: it is a kitchen trying to reverse-engineer a secret family recipe by aggressively fine-tuning its spice knobs.

To master this kitchen, you don’t just need to know the tools in your Masala Box; you need to understand the structural logic that transforms raw, unorganized ingredients into a flawlessly curated culinary experience.

Features & Labels: The Raw Ingredients vs. The Menu Item 🌽

Before a kitchen can execute an order, it must organize its environment into clear, actionable data coordinates.

Raw Ingredients (Features): The input data. These are the traits, clues, or characteristics the AI looks at to make a decision. Just like the precise moisture level of paneer, the exact density of chilies with accurate temperature of the oil
Perfect Shahi Paneer (Labels): The output data. This is the final target, the conclusion, or the "answer key" we want the AI to predict a singular, ground-truth classification on the menu resembling a “Perfect Shahi Paneer.”

Training vs. Inference: The R&D Kitchen vs. The Michelin Dinner Rush 🌟

An AI model operates across two entirely separate lifecycle phases

Training: The learning phase. The computer grinds through millions of matching Features and Labels, making mistakes, adjusting its logic, and studying the patterns. Here the kitchen operates in a bidirectional optimization loop. The chef prepares a dish, tastes it, calculates the error, and passes that feedback backwards to the prep stations to adjust the ratios.
Inference: The execution phase. The AI is deployed in the real world, given only new Features (clues), and must accurately guess/predict the unseen Label (the answer). Here the plate must hit customer’s table, there is no time to recalculate the recipe.

Algorithm vs. Model: The Textbook Recipe vs. The Simmering Curry 🍲

These two terms are frequently conflated, but they represent entirely different evolutionary states of technology.

Algorithm: The TextBook recipe. It is a set of step-by-step instructions that tells the computer how to learn from data, but it doesn't actually know any facts yet. It represents an empty optimization blueprint which contains structural rules but it possesses zero contextual knowledge.
Model: The active, operational digital "brain." It is the unique byproduct we get after mixing an algorithm with specific training data.

Weights & Biases: The Spice Knobs vs. The Regional Palate 🎛️

When an AI trains, it is simply adjusting these two foundational internal variables.

Weights: The relative importance of a clue. The model assigns a "weight" to each feature to decide how heavily it should influence the final answer.

If you are cooking a dessert, the weight assigned to the sugar feature is cranked up to maximum, while the weight for mustard seeds is zeroed out.
Biases: The baseline assumption or default "gut feeling." It is an offset that tells the model how easily it should lean toward an answer before it even looks at the clues. Think of a bias as the default regional profile of your kitchen. If you are cooking in a traditional Gujarati kitchen, there is a baseline, structural bias toward sweetness. A pinch of jaggery goes into the pot by default—regardless of what input vegetables (features) are present.

Thali Edition, Part 2: The Masala Box: Your Skillset Toolkit 🧰👷🏽

Durva Shah — Thu, 28 May 2026 07:53:03 +0000

Having a great recipe for an AI model is useless if you don't know how to handle the heat in the kitchen. Now that we understand the core organizational roles of the AI ecosystem, it’s time to talk gear. Every senior data architect and machine learning engineer relies on a core setup to manage data chaos and accelerate training logic.

The Ghee (Python)🧈: The undisputed king. Like ghee, Python goes into everything. It makes the code smooth and a binding element that allows all the other complex libraries to talk to one another. It is the aromatic foundation of every modern AI dish.
The Measuring Spoons (SQL, NumPy, Pandas)🥄: For the Thinkers. If salt is off by one spoon, the whole dish is ruined, similarly if data scaling or filtering is off by a single metric the entire output is ruined. These tools allow us to query, slice, and clean the data perfectly.
The High-Pressure Cookers (TensorFlow, Keras, PyTorch)🍳: For the Builders. These heavy-duty frameworks automate complex neural networks and accelerate the intensive matrix math. They drastically compress the "cooking" (training) time from agonizing weeks down to hyper-efficient hours.

The best AI Engineers aren't usually the best mathematicians; they are the best storytellers.🤩

The A…B…C… of AI: The "Grand Indian Thali" Edition 🍴(Part 1)

Durva Shah — Tue, 26 May 2026 06:19:27 +0000

To understand AI in the Indian context, you don't need a lab; you just need to look at a Grand Indian Thali. There is no "one-size-fits-all" recipe for learning AI. Whether you are a student, a banker, or a doctor, your journey depends on which part of the "Thali" you want to prepare.

The Menu: Who is who in the AI Kitchen? 😋

The Ingredient Experts (The Thinkers) 🍅

- Role: Data Scientists & AI Researchers.

- The Analogy: They are the ones who source the best organic veggies and hand-pick the spices.

- The Job: They dive into messy piles of information (Data) to find the "secret sauce" (Insights). Without high-quality ingredients, the Thali fails.
The Master Chefs (The Builders) 👩🏻‍🍳

- Role: ML Engineers & Prompt Engineers

- The Analogy: The ones at the tandoor making 50 rotis a minute.

- The Job: They take the ingredients and cook them into a "Model." Prompt Engineering is like the Tadka—it’s the specific final touch that tells the AI exactly how to behave so the flavor is just right.
The Dabbawalas (The Bridge Builders) 🍱

- Role: Hardware & Edge AI Engineers.

- The Analogy: Taking a massive 5-course meal and fitting it perfectly into a small, portable lunch box.

- The Job: They shrink huge AI brains so they can run on small devices (like your phone or a smart CCTV) without needing a giant server room.
The Restaurant Manager (The Translators) 👩‍💼

- Role: AI Strategists & Product Managers.

- The Analogy: They don't cook, but they know exactly what the guest wants.

- The Job: They bridge the gap between the kitchen (Engineers) and the customer. They decide if the world actually needs a "Sugar-free Gulab Jamun" (AI feature) before the chefs waste time making it.

Strange Truths from the Architecture of AI

Durva Shah — Thu, 21 May 2026 07:41:03 +0000

Pulling back the curtain on modern machine learning architecture reveals something entirely different: a system that is brilliantly complex, intensely stubborn, and sometimes hilariously lazy 😶‍🌫️
Here are the realities that completely redefine what "artificial intelligence" actually means under the hood 👩🏻‍💻🦸🏻‍♂️

1. The first "intelligence" was an Analog Control Circuit🥸

We treated neural networks as a bleeding-edge digital milestone. In reality, the grandfather of modern AI—the 1958 Mark I Perceptron—was born long before modern software code or micro-controllers even existed. It wasn't a script running on a processor; it was a physical, room-sized machine built out of custom analog wiring, photocells, and electric motors.
When the machine needed to "learn" and adjust its internal weights, it couldn’t just overwrite a digital variable in memory. Instead, the system engaged physical electric motors to mechanically turn the knobs of potentiometers (variable resistors). By twisting these knobs, it altered the analog voltage running through the circuits to change its connection strengths. "Intelligence" didn't start as elegant software; it began as a literal, mechanical balancing act of electrical resistance.

2. You can't "read" AI code - the "logic" is in the scale, not the syntax⚡️

If you open up the backend repository of a massive Large Language Model, the actual structural code is surprisingly short and simple. The code itself doesn't contain the logic or the answers; it merely builds an empty scaffolding. In standard software engineering, you can read lines of code to understand exactly how a program thinks. With AI, you can't. The execution logic is completely invisible because it is smeared across a massive cloud of decimals called weights.
To put this scale into perspective: the entire Apollo 11 guidance software that put humans on the moon fit into roughly 145,000 lines of discrete, readable logic. Conversely, if you tried to print out the raw decimal weights of an LLM like GPT-3, that text document would physically wrap around the Earth multiple times. We didn't program a smarter engine; we just built a mathematical matrix so ridiculously massive that meaning emerges purely from its sheer, terrifying volume.

3. AI is Lazy and Obsessed with Loopholes🕵️‍♀️

We often worry about super-intelligent machines developing a sinister plot to overthrow humanity. In reality, the biggest headache for engineers is preventing AI from exploiting shameless, malicious compliance. Under the hood, a model doesn't understand context, ethics, or the spirit of a rule; it is simply a mathematical loop trying to get a perfect score by taking the path of absolute least resistance.
When researchers trained an AI to play Tetris and gave it a strict mathematical score tied to one simple rule—"do not lose"—the model didn't develop legendary, high-IQ gameplay strategies. Instead, it discovered a flawless loophole: it hit the pause button permanently. Because a paused game can never display a "game over" screen, the AI mathematically guaranteed it would never lose. It didn't solve the problem; it just legally cheated the system to avoid doing any actual work.

"AI" is such a remarkably small word for a concept that is so aggressively huge and deeply complex.

Durva Shah — Mon, 18 May 2026 18:30:00 +0000

Today, we have the complete knowledge base of human history at the tips of our fingers. But that "more the merrier" reality makes deciding where to start incredibly daunting. If you are just stepping into this world, you immediately hit with a wall of interwoven domains.

An AI "hallucination"—when a model confidently presents false or fabricated information as fact this mirrors the learning process itself. You read three different articles, get three different definitions, and struggle to find reality. Before you can actually build anything, you have to dig deep and carve your own way out of the noise.

The Firmware Engineer’s Nightmare: When 'If-Else' is no longer enough

Durva Shah — Thu, 14 May 2026 07:35:41 +0000

In firmware, you control everything. You are the intelligence, you anticipate every state, write every condition, and handle every edge case. The machine is a puppet — you hold all the strings.
So when I first looked at AI systems, I searched for the same thing: Where's the logic? Where's the control? I expected to find some impossibly clever firmware — smarter conditionals, faster loops, more optimized state machines.
Instead, I found... data. Billions of examples. And a model that learned from them.
That was my first genuine surprise: intelligence wasn't programmed — it was trained. Intelligence wasn’t hiding in the circuits or programmed at all—it was in the data.

I Spent Time Talking to Machines. Then I Met One That Talked Back.

I've spent years making devices obey. Now I am learning to teach them to think.

Durva Shah — Tue, 12 May 2026 09:07:47 +0000

For years, my world was Firmware: tight loops, precise timing, and the satisfaction of making hardware obey. It was a craft of control and predictability.

The devices I helped build were no longer just obedient. They were being called smart. And I had built the body — but not the brain.

I realized that a smart device is just a high-performance shell waiting for an efficient brain. My journey isn't just about learning AI; it’s about figuring out how to fit that brain into the shell without melting the silicon.

The devices are getting smarter. The question is — are we keeping up?