Blog

Here are some of my recent articles on Medium. You can find more on my Medium profile.

• 
  Gemma from Scratch: Mastering LLM Implementation with JAX and Flax

  Guide on porting Gemma to JAX/Flax, implementing core components like RoPE and GQA to leverage XLA's performance on TPUs and GPUs.

  Dec 20, 2025

• 
  Your AI from Scratch: Building a Customised LLM

  Foundational tutorial implementing a Transformer model from zero using PyTorch, demystifying attention, normalization, and embeddings for deep architectural understanding.

  Dec 20, 2025

• 
  Google Colab & VS Code Extension

  Workflow guide for connecting local VS Code to Colab's cloud GPUs/TPUs, enhancing productivity with IDE features while using remote hardware.

  Nov 13, 2025

• 
  Fine-tune Gemma-3–270M for Financial Sentiment Analysis

  Demonstrates high-performance financial sentiment analysis using the ultra-lightweight Gemma-3 270M model, highlighting efficiency for resource-constrained on-device AI applications.

  Sep 11, 2025

• 
  Fine-Tuning Gemma 3 1B for Function Calling: A Step-by-Step Guide

  Tutorial on enabling agentic capabilities in Gemma 3 1B via LoRA fine-tuning for structured tool-use and reliable function calling.

  Jun 25, 2025

• 
  Building agents with Gemini 2.5 Pro for the HF Agents Course

  Evaluates Gemini 2.5 Pro on the GAIA benchmark, exploring advanced reasoning and long-context windows for complex, multi-step agentic tasks.

  Jun 18, 2025

• 
  Learning from Kaggle Competitions using Gemini 2.5: AI Mathematical Olympiad — Progress Prize 2

  Strategic overview of solving National Olympiad-level math problems using Gemini 2.5, focusing on majority voting and prefix caching techniques.

  May 18, 2025

• 
  Fine-Tuning Gemma 3 1B-IT for Financial Sentiment Analysis: A Step-by-Step Guide

  Step-by-step guide for specialized financial sentiment analysis using PEFT/LoRA on the Gemma 3 1B-IT model for domain-specific accuracy.

  Mar 25, 2024

• 
  Training for Reasoning with GRPO — part II (a step by step explanation)

  Technical deep-dive into implementing Group Relative Policy Optimization (GRPO) with Gemma 2 2B-IT to teach models explicit reasoning traces.

  Mar 7, 2024

• 
  A Personal Assistant for knowledge management based on Gemini on Gemini 2.0 and Vertex AI

  Building a custom AI assistant using Gemini 2.0 Flash to organize and summarize personal notes and URLs into structured knowledge.

  Feb 26, 2024

• 
  Training for Reasoning with GRPO — part I ( project overview & results)

  Research overview of applying GRPO reinforcement learning to Gemma 2 2B-IT, enabling emergent reasoning capabilities in small language models.

  Feb 21, 2024

• 
  Gemma 2 2B learns how to tutor in AI/ML

  Shows how to create a specialized AI tutor by fine-tuning Gemma 2 2B on synthetic data generated from technical documentation.

  Sep 25, 2023

• 
  Data Science AI Assistant with Gemma 2b-it: a RAG 101

  Bare-metal implementation of Retrieval-Augmented Generation (RAG) using ScaNN and GTE-large to ground Gemma 2B-IT in factual knowledge.

  May 9, 2023

• 
  Sherlock Holmes Q&A Enhanced with Gemma 2b-it Fine-Tuning

  Demonstrates baking niche domain knowledge into Gemma 2B-IT's weights through fine-tuning, serving as an alternative to RAG for specific topics.

  Apr 2, 2023

• 
  ML Olympiads: detect hallucinations in LLMs with Google Gemini

  Uses Google Gemini as an automated evaluator to detect factual inconsistencies and hallucinations in other LLMs for improved reliability.

  Mar 20, 2023

• 
  Higher performance with Gemma

  Introductory overview of Google's Gemma models, showcasing their benchmark-surpassing performance and suitability for on-device AI and local development.

  Feb 29, 2023

• 
  Fine-tuning a large language model on Kaggle Notebooks for solving real-world tasks — part 5: saving your work for reusing

  Concluding guide on merging LoRA weights, managing GPU memory, and saving models for production deployment after fine-tuning on Kaggle.

  Feb 15, 2023

• 
  Fine-tuning a large language model on Kaggle Notebooks for solving real-world tasks — part 4 (LLama 2 strikes back)

  Advanced optimization techniques for conversational fine-tuning, focusing on chat formats and strategic layer selection for better parameter efficiency.

  Dec 15, 2022

• 
  Fine-tuning a large language model on Kaggle Notebooks for solving real-world tasks — part 3

  Adapts the fine-tuning workflow to Mistral 7B and Phi-2, demonstrating high performance on financial tasks with smaller, efficient models.

  Nov 15, 2022

• 
  Fine-tuning a large language model on Kaggle Notebooks for solving real-world tasks — part 2

  Practical tutorial on the end-to-end pipeline for financial sentiment analysis using Llama 2 7B on Kaggle's free GPU resources.

  Oct 15, 2022

• 
  Finetuning a large language model on Kaggle Notebooks for solving real-world tasks — part 1

  Theoretical foundation for LLM fine-tuning, explaining QLoRA and PEFT as solutions for adapting large models with limited hardware.

  Sep 15, 2022