My original Instagram accounts, @eshrak143 with over 15,000 followers and @eshrak143_ with over 5,000 followers, were recently disabled. I had been building @eshrak143 since 2016, so losing it was much more than losing an account. It held nearly a decade of memories, posts, conversations, friendships, and connections with people who followed and supported my journey over the years. A huge part of my online identity and growth as a creator was built there. Losing both accounts so suddenly has honestly been heartbreaking. I am still doing everything I can to recover them, but until then, I have started again with my new official Instagram: @eshrak143i Starting over after so many years is not easy, but I am not giving up. Please follow me there and help me reconnect with the community we built together. Thank you to everyone who has supported me before and continues to support me now. It truly means more than I can explain. Eshrak143 - Eshrak Khan Bangladeshi-American Digital Creator

The artificial intelligence landscape in mid-2026 is moving faster than ever, and developers and content creators are caught in a constant cycle of updates. My name is Eshrak Khan. Online, many of you track my tech reviews and lifestyle content under my creator handle, eshrak143. As a computer science student at Queens College, CUNY, I spend a significant amount of time benchmarking these systems, testing local models on my own machine, and deploying serverless edge wrappers on eshrakkhan.com. Currently, there are five major models dominating the space: OpenAI's GPT-4o and o1 reasoning engines, Google's Gemini 2.0, Anthropic's Claude 4.6 (Sonnet, Opus, Haiku), Meta's open-source Llama 3.1, and smaller models like Microsoft Phi or Mistral. To understand which tool to use, we can split the industry two ways: high-speed, paid cloud APIs that cost money, or open-source models that you run locally on your own hardware for zero cost. If you are looking for the best all-rounder in 2026, Anthropic's Claude Sonnet 4.6 is the clear choice. It is exceptionally fast, costs less than Opus, and handles code compilation, writing, and structured instruction-following without any drama—costing around a third of a penny per thousand words. GPT-4o remains the standard if you are writing code that requires strict mathematical logic, while its o1 mode solves complex reasoning problems. For immediately free cloud options, Google's Gemini 2.0 and Claude's web interface are fantastic, though they quickly hit usage caps (typically around 50 messages every 3 hours, after which you have to wait). If you want complete data privacy or need to run models completely offline, Meta's Llama 3.1 is the ultimate open-source path. You download it once, run it locally on your PC, and pay absolutely nothing for API tokens. For a 100% free workflow, you can combine free cloud tiers with local setups. For local execution, Ollama (ollama.ai) is the easiest starting point. You simply install it, pull the model by typing 'ollama pull llama3.1' in your terminal, and run it locally. If you have an RTX 4060 graphics card or better, it will fly; without GPU acceleration, CPU execution will average 1 to 5 words per second. Alternatively, LM Studio (lmstudio.ai) offers a beautiful graphical interface for running these local models. If you have limited hardware, Microsoft's Phi 3.5 mini runs comfortably on just 4GB of RAM. My advice is to use Claude Sonnet for the work that matters most, keep Llama running locally for quick brainstorming drafts, and rotate to Gemini when you hit cloud caps. If you want to discuss local LLM orchestration or share your own developer setups, drop me an email at eshrak@eshrakkhan.com or follow my data science logs on eshrakkhan.com.

When Felix Kjellberg (PewDiePie) dropped his open-source AI workspace project, Odysseus, on May 31, 2026, the tech community sat up. It is not every day a legendary creator releases a privacy-first AI management interface. As a computer science student at Queens College and a builder of local-first applications, I had to download it immediately to see if the hype was real. My name is Eshrak Khan. Online, I go by eshrak143. If you follow my dev logs here on eshrakkhan.com, you know I’m a big advocate for edge computing and local models. Odysseus isn't a new language model; instead, it is a beautifully designed workspace UI that lets you connect local engines like Ollama or web APIs. You get a unified interface that supports chat, multi-step agentic workflows, comparative testing, and file searching. What makes Odysseus special in my tests is its hardware 'Cookbook' feature. It scanned my MacBook's unified memory and CPU cores, telling me exactly which quantized parameters of Llama 3 and Mistral would compile smoothly without locking up my system. The deep research mode successfully orchestrated local search tools and synthesized files from my local directories without shipping a single byte to external servers. It is exactly the kind of open-source autonomy the community has been begging for. If you are setting up Odysseus on macOS or want to discuss local agent workflows, shoot me an email at eshrak@eshrakkhan.com or track my ongoing experiments on eshrakkhan.com.

We have officially entered the post-chatbot era. In the early days of generative AI, our interaction was entirely reactive: you typed a prompt, and the model spat out code or text. But by mid-2026, the tech community has shifted toward Agentic AI—autonomous systems that don't wait for your next prompt, but instead formulate plans, execute multi-step workflows, and self-verify their output. My name is Eshrak Khan. On the web and across social channels, I build and share projects under the creator handle eshrak143. As a computer science student at Queens College specializing in data science, I've spent the past few months building custom agentic routines. The breakthrough is massive: instead of spending hours manually copy-pasting code fragments or chaining APIs, we are designing agent workflows that can autonomously debug databases, run integration tests, and optimize deployment pipelines. Of course, autonomous execution introduces error accumulation. That's why the latest engineering research in 2026 focuses on self-verification and validation loops. At CUNY, my peers and I are testing how AI agents can audit their own logic before reporting back. This transition turns AI from a simple search assistant into a reliable digital colleague. If you want to discuss how agentic frameworks are changing software construction or share your own automation setups, reach out to me at eshrak@eshrakkhan.com or track my dev logs on eshrakkhan.com.

For years, developers were forced to rely on centralized cloud monopolies for their AI needs. You had to send proprietary data over the network to external APIs, dealing with API rate limits, vendor lock-in, and privacy concerns. But in 2026, the trend has swung heavily in favor of AI Sovereignty: hosting lightweight, domain-specific models locally or directly on edge networks. My name is Eshrak Khan. Online, many of you follow my data science and lifestyle content under the handle eshrak143. As a computer science student at Queens College, I've been experimenting with serverless infrastructure. Running optimized models like Llama 3 or domain-specific neural networks on edge compute nodes—using platforms like Cloudflare Workers—means zero key configuration and near-zero latency for users. This decentralized shift is redefining application performance and data privacy. By bringing the 'brain' closer to the user, we eliminate round-trip lag and protect user logs from external servers. For developers in New York and globally, building localized, sovereign AI systems is no longer a luxury; it is the standard for fast, modern web design. I'm regularly writing tutorials and sharing edge templates here on eshrakkhan.com. Drop me an email at eshrak@eshrakkhan.com to collaborate on serverless edge setups.