• AI/ML - Articles

    Hosting a SLM Qwen 2.5 on Raspberry Pi 2

    December 8, 2025 - By 

    This article is a write up of my experience on hosting Qwen 2.5 the 0.5B model in Raspberry Pi 2 using Llama.cpp. Qwen 2.5 is one of the small SLM with 0.5B parameters so a small development board like Raspberry Pi can hold it. RPi 2 Model B comes with 900Mhz speed and only 1GB of memory. But to be honest, setting up the project might take 1-2 hours, and the prompt execution is only some 1-2 tokens per second. So you need to be patient. Let us begin! Step 0: Pick up the Raspberry Pi 2 from the attic.…

  • AI/ML

    Microsoft Fara-8B learning experience

    December 7, 2025 - By 

    Over the weekend, I was trying do a learning project using Microsoft’s latest Microsoft Fara-7B model, which is a “Computer Use Agent”. Microsoft Fara-7B is a small, efficient, open-weight AI model (with 7 billion parameters) designed to act as a Computer Use Agent (CUA), allowing it to perform tasks on a computer by visually understanding the screen (aka screenshots) and using mouse/keyboard actions (clicks, typing, scrolling) to automate web tasks like booking travel, shopping, or filling forms, offering speed, privacy (runs locally on devices), and lower cost compared to larger models. Currently the model is still in experimental mode, and…

  • AI/ML - Articles

    Strategy for adding AI to the existing application

    December 4, 2025 - By 

    How should a technical leader respond when a customer asks to add AI to an existing application? The answer requires structure and clear thinking. 1. First, clarify the Actual ProblemNever assume that AI is the right solution. I would start by understanding the business objective. Many requests framed as AI needs turn out to be workflow issues, reporting gaps, or rule based automation opportunities. Accurate problem definition prevents unnecessary complexity. 2. Evaluate Data/App ReadinessAI depends completely on data quality. Assess what data exists, how clean it is, and whether privacy or compliance concerns limit its use. If data foundations are…

  • AI/ML - Articles

    The “Stochastic Parrot” problem and why it still matters in AI system design

    December 4, 2025 - By 

    The term stochastic parrot was introduced in a 2021 paper by Bender, Gebru, and colleagues (ref: Wikipedia). It highlights a fundamental limitation of large language models. These systems generate text by predicting the next token based on statistical patterns. They do not possess grounded understanding of the world. This can lead to convincing output that is incorrect, biased, or superficial. What the metaphor captures is simple: the probabilistic, statistically driven nature of these models. Parrot evokes an entity that mimics language without real understanding. The critique is not about style. It is about reliability. When a model draws from vast…

  • AI/ML

    Recursive self-improvement of AI prompts

    November 29, 2025 - By 

    Large Language Models are probabilistic. They predict the next most likely word. When you ask them to “critique,” you populate their context window with high quality reasoning and negative constraints (eg. what not to do). The final generation is then statistically more likely to follow that higher standard because the logic is now part of the immediate conversation history. Try this: Draft: Ask for your content as usual. “Write a cold email to a potential client about our new web design services.” Critique: Dont just ask for a better version. Ask the AI to analyze its draft against specific criteria.…

  • AI/ML - Code

    Code snippet: Hello World agent using Microsoft Agent Framework

    November 22, 2025 - By 

    This code snippet demonstrates a sample code which uses Azure OpenAI endpoint to execute an LLM call. # pip install agent-framework python-dotenv import asyncio import os from dotenv import load_dotenv from agent_framework.azure import AzureOpenAIChatClient # Load environment variables from .env file load_dotenv() api_key = os.getenv("AZURE_OPENAI_API_KEY") deployment_name = os.getenv("AZURE_OPENAI_DEPLOYMENT") endpoint = os.getenv("AZURE_OPENAI_ENDPOINT") api_version = os.getenv("AZURE_OPENAI_API_VERSION") agent = AzureOpenAIChatClient( endpoint=endpoint, api_key = api_key, deployment_name=deployment_name, api_version=api_version ).create_agent( instructions="You are a poet", name="Poet" ) async def main(): result = await agent.run("Write a two liner poem on nature") print(result.text) asyncio.run(main()) .env file sample AZURE_OPENAI_API_KEY={paste your api key} AZURE_OPENAI_ENDPOINT=https://{your enpoint}.openai.azure.com/ AZURE_OPENAI_DEPLOYMENT=o4-mini AZURE_OPENAI_API_VERSION=2024-12-01-preview

  • AI/ML

    Completed: Course on Semantic Caching for AI Agents by deeplearning.ai

    November 21, 2025 - By 

    Most of us are in a transition phase from AI prototypes to production systems. Many frameworks that appeared impressive on demo servers have failed badly in real production environments. It is important to consider all architectural pillars and aspects during the design stage. Delaying these decisions only adds time and cost later. Agentic/AI consumes tokens, and token usage directly translates to monetary cost. This course offers a clear explanation of AI/agent caching techniques and shows how to evaluate the effectiveness of different caching strategies. Attend the course here: https://www.deeplearning.ai/short-courses/semantic-caching-for-ai-agents/

  • Code

    CaskIA — Kid-friendly, minimal AI Chat

    November 17, 2025 - By 

    My 3rd-grade kid has been using modern chat assistants for his hobbies a lot lately. Today he said he wanted his own AI chat bot. He tried building one with GitHub Copilot, but the chain-of-thought prompting pushed him into creating a search engine style chat since he did not realize he needed to connect an actual LLM to make it intelligent. I stepped in and helped him build a simple app using the free version of GitHub Copilot, and also explained how it works. It took about 30 minutes to put everything together. I am sharing it publicly so others…