Romee Panchal
@romitheguru
ML Engineer | Writer, thinker, and lifelong learner. Subscribe to my newsletter: https://romeepanchal.substack.com/
Don't read to question it. Don't read to agree with it. Read to understand it.
Even with cutting-edge ML models, garbage in = garbage out! Messy data tanks performance every time. Pro tip: Master cleaning & prepping your data—it’s a valuable data science skill! #DataScience #MachineLearning
Instead of hype, focus on practice: Learn how to build agents with modular tools, test in real scenarios, and iterate relentlessly. Value? Autonomous efficiency, scalable problem-solving, and human creativity amplified. #AIAgents #AI
Good stuff. I tried to generate a few scripts. Worked well.
>>> Qwen3-Coder is here! ✅ We’re releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date. This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation. It achieves…
I asked Claude to be skeptical, question me for clarity, and even challenge me with real questions. It works well. Now, every time I ask something, it poses questions to clarify until the full context is clear. I think it is better this way.
There is big between an MVP and product. Don't spend too much on prototype Build it, pitch it, and gather feedback to refine and build your product.
How to ensure your data is representative for a model: follow these steps: 1. Understand the Target Population: Define the real-world population or scenarios the model will address to set a benchmark for representativeness. 2. Compare Data Distribution: Check if the data’s…
How To Read AI Research Papers Effectively? Method 1: Ask ChatGPT to explain it to you like you are 10. Method 2: Follow this video: youtu.be/K6Wui3mn-uI
Despite advanced machine learning models, the key challenge is data. Messy data leads to poor performance, even with the best models. Learn to clean and prepare data for your problem—it's one of the most valuable skills in data science.
Ollama vs. vLLM: both tools simplify running LLMs. Key differences: - Ollama prioritises simplicity, local execution, and accessibility, supporting CPU and GPU with minimal setup. - vLLM focuses on performance, scalability, and GPU optimisation for high-concurrency workloads.
From Word Vectors to Reasoning Models: The Engineering Evolution of NLP medium.com/p/from-word-ve…
Either some people lack basic comprehension skills, or they intentionally misinterpret things. I say, "Don't spend too much time on theory". What they interpret as "Oh, so you are saying you don't need to learn theory". These are two different things.
Microsoft recently released the Phi-4-mini-flash-reasoning model. This model is optimised for edge devices and mobile applications, offering efficient reasoning with a compact architecture. Read here: azure.microsoft.com/en-us/blog/rea…
Unless you're interested in ML research, you don't need to spend too much time exploring the math and architectures of every ML and LLM algorithm. Focus on learning how to build real-world ML systems, the trade-offs of different algorithms, and their engineering side.
Modern NLP development follows a pattern where a breakthrough emerges from solving concrete computational bottlenecks in existing systems. A wrote an article discussing the progress in NLP. romeepanchal.substack.com/p/from-word-ve…