LLM steps

The rapid evolution of Large Language Models (LLMs) has transformed how we interact with technology. From writing code to offering medical insights, these systems feel almost human. But behind the "magic" lies a rigorous, multi-stage lifecycle. 

Based on the expert roadmap provided by Rocky Bhatia, here is a breakdown of how an LLM is built, refined, and maintained.

Stage 1: The Foundation – Data & Architecture

Before a model can "think," it must be fed. This initial phase focuses on gathering the raw materials of intelligence.

Stage 2: Training – Learning to Predict

With the architecture set, the model begins its "education."

Stage 3: Alignment – Fine-Tuning for Quality

A model that knows everything is useless if it isn't helpful or safe. This stage "aligns" the model with human values.

Stage 4: Deployment & Interaction

Once the model is accurate and safe, it is ready for the real world.

Stage 5: Evolution – Performance & Continuous Learning

The journey doesn't end at deployment. Language and information are always changing.

Enterprise AI strategy

 

This image perfectly captures the "AI Paradox" that many enterprise leaders are currently facing. While the board is excited about the shiny "AI Strategy" peak, the engineering teams are deep underwater dealing with the messy reality of technical debt.

Every executive team wants to "move fast" on AI. They see the summit—generative models, automated insights, and transformative growth. But beneath the surface of a sleek AI Strategy lies a massive, cold reality: the technical debt of the last twenty years.

If you try to build a high-speed AI engine on a foundation of "spaghetti" code and siloed data, the iceberg won't just slow you down—it will sink the project.

To move from the "Excited Executive" boat to the "Successful Implementation" submarine, teams need to stop treating AI as a separate project and start treating it as the ultimate reason to clean house.

• Audit Before You Automate: Document your legacy logic before asking an AI to refactor it.

• Data Hygiene is AI Strategy: If your data is messy, your AI will just be "fast and wrong."

• Empower the "Foundations" Team: Give the engineers in the middle boat the resources they need to fix the base. You can't reach the peak if the bottom is crumbling.

2026 Case Studies

Theoretical knowledge of databases and load balancers is a start but seeing how they are applied under extreme pressure is where the real learning happens. These case studies teach you about:

• Scalability: How to grow from 1,000 to 1 billion users.

• Availability: Ensuring services never go down, even during updates.

• Latency: Reducing the milliseconds between a user's click and the system's response.

1. YouTube & MySQL (The Power of Sharding)

Most people assume a giant like YouTube would abandon traditional relational databases for something "fancier." However, YouTube famously scaled MySQL by using Vitess, a database clustering system.

• The Lesson: You don't always need a NoSQL database to scale; horizontal sharding can allow traditional SQL databases to handle massive throughput.

2. Uber’s Geospatial Indexing

Uber processes over 1 million requests per second to match riders with drivers. This isn't just a simple query; it requires calculating distances in real-time.

• The Lesson: Uber uses hexagonal hierarchical spatial indexing (H3). By dividing the world into hexagons, they can quickly calculate "proximity" without taxing the CPU with complex geometry.

3. Meta’s Serverless Scale

Handling 11.5 million serverless function calls per second is an exercise in extreme orchestration. Meta (Facebook) has to manage resource allocation and "cold starts" at a scale most companies will never see.

• The Lesson: Serverless isn't just for small hobby projects; with the right infrastructure, it can power the world's largest social network.

System design is less about finding the "perfect" tool and more about understanding trade-offs. Whether it's choosing between the consistency of Google Docs or the massive throughput of Kafka, these 12 examples provide the blueprint for modern engineering.

AI Journey 2025 to 2026

As 2025 draws to a close, the AI landscape has shifted from "chatbots as a novelty" to "AI as an omnipresent teammate." We’ve moved beyond asking if AI can write an email to asking which model should run our entire department’s workflow.

Here is a year-end wrap-up of the "Big Three"—ChatGPT, Gemini, and Claude—and how they’ve redefined themselves in 2025.

Feature	ChatGPT (GPT-5.1)	Gemini (3 Pro)	Claude (4.5 Sonnet)
Best For	Daily Versatility	Multimodal/Big Docs	Coding & Nuanced Writing
Context Window	~128K - 1M (varies)	1M - 2M Tokens	200K - 1M (beta)
Key Innovation	Thinking Mode (Adaptive)	Computer Use/Native Video	Claude Code / Agentic Flow
Personality	Direct & Intuitive	Analytical & Integrated	Thoughtful & Cautious

Looking Toward 2026: The "Invisible" AI

As we flip the calendar, the trend is clear: AI is becoming ambient. We are moving away from dedicated apps and toward a world where AI quietly operates in the background—summarizing our meetings, organizing our files, and predicting our needs before we even type a prompt.

The bubble hasn't burst; it has simply soaked into the fabric of how we work.

Final Thought: In 2023, we were amazed it could talk. In 2024, we were amazed it could see. In 2025, we’ve finally started letting it do.

Happy New Year 2026 to all of you!

 

Code to Compost

As we close out 2025, I’ve been reflecting on the diverse paths our lives take. This week’s blog is a tribute to my UnderGrad classmate, Partha, whose journey over the last decade serves as a profound masterclass in purpose and resilience.

Life Transformation

After twenty years in the software industry, Partha made a choice many of us contemplate but few dare to execute: he traded the corporate office for the open field. Today, he is a seasoned natural farmer, leading the way in sustainable agriculture at the Bhoomi Network’s Farmer’s Center in Tiruvannamalai, India.

The Courage to Pivot

Spending time with Partha at his farm was a mesmerizing experience. Witnessing his transition from two decades of IT logic to the intuitive, rhythmic cycles of nature was deeply humbling. While many of us talk about "getting back to our roots," Partha has spent the last ten years living that reality—transforming rough, challenging land into a thriving ecosystem.

My visit left me with a few powerful realizations:

• The True Value of Nature: Seeing the interdependence of soil, water, and weather firsthand.
  
  
• Grit and Commitment: The physical and mental stamina required to build a farm from scratch is immense.
  
  
• A Student’s Mindset: Partha’s success stems from his passion for asking the right questions and his willingness to keep learning from the land.

A Hub for Ecological Change: The Vivasayi Maiyam

Partha’s work at the Bhoomi Network's Farmer’s Center (Vivasayi Maiyam) in Tamil Nadu is more than just farming; it is an educational movement. The center serves as a bridge for those looking to transition from conventional methods to "nature-imitated" practices through several structured programs:

1. Volunteering (7+ Days): A hands-on entry point for those over 18. Participants engage in farm work and group discussions under a unique "Gift Culture" model (suggested ₹500/day).
2. Village Immersion (3–5 Days): A program designed for urban professionals to experience rural social realities by living and working alongside village communities.
3. Holistic Internships (1 Month+): For those seeking a deeper, comprehensive understanding of ecological systems.
4. Model Farm Creation (6 Months+): A rigorous, experiential program where participants manage a one-acre plot, learning everything from seed banking to animal integration.

Final Thoughts

We often think of success in terms of vertical growth within an industry. Partha has shown me that success can also be horizontal expanding our impact across different sectors and aligning our work with the health of the planet.

I am incredibly proud of my friend. His journey reminds us that it is never too late to redefine who we are and what we contribute to the world.

TOON usage

In the rapidly advancing landscape of AI, TOON (Token-Oriented Object Notation) has emerged as a specialized alternative to JSON. It is designed specifically to optimize how Large Language Models (LLMs) like GPT-4, Gemini, and Claude process data.

Top 3 factors to leverage TOON

1. Drastic Cost Savings

AI models do not read text character-by-character; they process "tokens." In standard JSON, every brace { }, quote ", and comma , consumes tokens.

• Token Reduction: TOON can reduce token usage by 30% to 60% compared to JSON.

• Lower Bills: Since most AI providers charge per token, switching to TOON can cut your API costs nearly in half for the exact same data.

2. Expanded "Memory" (Context Window)

Every AI model has a limit on how much information it can "remember" at once (the context window).

• Efficiency: Because TOON is more compact, you can fit twice as much data into a single prompt.

• Better RAG: For Retrieval-Augmented Generation (RAG) systems, this means you can feed the AI more search results or documents without hitting the limit.

3. Improved Accuracy and Reasoning

JSON often forces the AI to navigate a sea of repetitive keys and brackets. TOON uses a tabular, header-based structure that aligns better with how models naturally "see" patterns.

• Reduced Hallucinations: With less syntactic noise, the AI can focus on the actual values, leading to more accurate data extraction and reasoning.

• Structure over Syntax: TOON explicitly declares array lengths and headers once, which helps the model stay "grounded" when reconstructing large tables.

AI Tools 2025 Pyramid

 

Level 1: Foundational Services

At the very base of the pyramid are the foundational AI services, which often work behind the scenes or provide core functionalities for more complex tasks.

• RecCloud (Video Translation Service): In an increasingly globalized world, video content is king. RecCloud helps break down language barriers by providing video translation services, making your content accessible to a wider international audience.

• Tidio (Customer Service): Excellent customer service is paramount for any business. Tidio leverages AI to power chatbots and streamline customer interactions, providing instant support and freeing up human agents for more complex issues.

• Zapier (Multi-platform Integration): While not strictly an "AI tool" in itself, Zapier is crucial for integrating various AI tools and automating workflows across different platforms. It acts as the connective tissue, allowing your AI applications to communicate seamlessly.

• Calendly (Schedule Management): AI-powered scheduling tools like Calendly take the hassle out of booking appointments. They can intelligently find the best times, send reminders, and manage your calendar, optimizing your time.

• PicWish (Image Generation): As we move towards more visual content, AI image generation tools like PicWish are becoming indispensable. They allow users to create stunning visuals from text prompts, saving time and resources on graphic design.

Level 2: Design Tools

The next level up focuses on AI tools that assist in design and creative endeavors, empowering both professionals and casual users to produce high-quality visuals and experiences.

• Midjourney (Image Generation): A powerful generative AI tool, Midjourney allows users to create incredibly detailed and artistic images from simple text descriptions, pushing the boundaries of visual creativity.

• Recraft (Image Generation): Similar to Midjourney, Recraft also specializes in AI-powered image generation, providing another avenue for creating unique visual content for various applications.

• Figma (Design Collaboration): While Figma is primarily a UI/UX design tool, its collaborative features can be enhanced with AI plugins, streamlining design processes and team communication.

• Canva (Design Tool): Canva has democratized design, and with its increasing integration of AI features, it makes graphic design accessible to everyone, offering smart recommendations and automated design elements.

Level 3: Writing Assistants

AI writing assistants are revolutionizing content creation, helping with everything from drafting emails to generating long-form articles.

• ChatGPT (Writing Assistant): A household name, ChatGPT is a versatile AI language model capable of generating human-like text, answering questions, summarizing information, and assisting with various writing tasks.

• Claude (Writing Assistant): Another advanced AI assistant, Claude excels at complex reasoning and generating coherent, thoughtful responses, making it ideal for in-depth content creation and analysis.

• Grok (Writing Assistant): Grok, with its real-time knowledge access, offers unique capabilities for generating timely and relevant content, especially useful for news and current affairs.

Level 4: Market Research

At this level, AI tools become more strategic, assisting with data analysis and market insights to inform business decisions.

• Perplexity (Market Research): Perplexity combines powerful search capabilities with AI summarization, making it an excellent tool for quickly gathering and understanding information for market research.

• Gemini (Market Research): Google's Gemini is a multimodal AI that can process and understand various types of information, making it a robust tool for comprehensive market analysis and trend identification.

Level 5: Code Editor

At the pinnacle of the pyramid lies the most specialized and advanced AI application for developers.

• Cursor (Code Editor): Cursor is an AI-powered code editor that assists developers by suggesting code, identifying errors, and even generating entire functions, significantly accelerating the coding process and improving code quality.