Soham Sharma
17 published articles across 10 categories
About
Soham Sharma contributes practical engineering and product insights on automation, AI, and modern web systems at Botmartz.
Contact: sohamnsharma@gmail.com
Published
17
Live articles
Categories
10
Covered topics
Latest Post
June 12, 2026
Last published date
Published Posts
Latest insights and articles by Soham Sharma
Research Paper Deep Dive: RoPE (Rotary Position Embeddings) — Better Position Information
Standard position embeddings are additive and have poor long-range generalization. RoPE embeds positions via rotation: multiply Q, K by rotation matrices. Enables 100K+ token context.
June 12, 2026
Language Model Architectures: Transformers, Attention, and the Path from GPT-1 to GPT-4
Modern LLMs are Transformers. Understand the evolution: self-attention, positional encoding, scaling laws, and how each architectural change improved performance.
June 12, 2026
Read article →AI Agent Fundamentals: Decision-Making Loops, Tools, and Agentic vs. Procedural Reasoning
Agents make autonomous decisions by reasoning, planning, and calling tools. Understand the perception-decision-action loop and when to use agents vs. deterministic workflows.
June 12, 2026
Read article →Mamba: State Space Models and the Alternative to Transformer Attention
Transformers require O(n²) attention. Mamba uses state space models for O(n) complexity with better scaling. Understand selective SSMs and why Mamba matches transformer quality at 1/5 the memory.
June 12, 2026
Read article →Research Paper Deep Dive: Flash Attention 2 — Optimizing Transformer Attention
Flash Attention achieves 2-4× speedup on attention by changing memory access patterns. Understand I/O complexity, tiling, and how to optimize matrix operations on GPUs.
June 8, 2026
Read article →Optimizer Comparison: SGD, Momentum, Adam, RMSprop, and When Each Shines
Different optimizers suit different problems. SGD is stable, Momentum accelerates, Adam is adaptive. Understand why each optimizer works and pick the right one.
June 8, 2026
Read article →Vision Transformers (ViT): Image Classification with Pure Transformers
Vision Transformers apply Transformers to image classification. Patch embeddings convert images to sequences, enabling the same architecture as NLP models.
June 8, 2026
Read article →
Custom Training Loops with GradientTape: Manual Forward and Backward Passes in TensorFlow
model.fit() hides the training loop. GradientTape exposes it. Use it when you need per-batch gradient manipulation, custom loss combinations, or training dynamics that Keras callbacks can't express.
June 3, 2026
Read article →
Rotary Positional Embeddings (RoPE): How It Works and Why It Beats Learned Embeddings
RoPE encodes position by rotating query and key vectors in complex space. It extrapolates beyond training length, transfers across fine-tuning, and adds zero parameters — here's the math.
June 3, 2026
Read article →
PyTorch Custom Dataset and DataLoader: __getitem__, __len__, collate_fn, and num_workers
DataLoader is more than a loop — it's a parallel data pipeline. Build a correct Dataset, write a proper collate_fn, and understand num_workers to eliminate training bottlenecks.
June 3, 2026
Read article →
Working with LLMs and Chat Models in LangChain: OpenAI, Anthropic, and Local Models via Ollama
LangChain wraps every LLM provider behind the same Runnable interface. Swap OpenAI for Claude or a local Llama model without changing a line of your chain logic.
June 2, 2026
Read article →Keras Sequential vs Functional vs Subclassing: When to Use Which API
Keras gives you three model-building APIs. Sequential is a dead end for anything non-trivial. Functional handles 90% of production architectures. Subclassing gives you full control when you need it.
April 27, 2026
Read article →PyTorch Autograd Internals: Computation Graphs, retain_graph, grad_fn Chain, and detach
Autograd is not magic — it's a directed acyclic graph of Function nodes. Understand how gradients flow, when retain_graph matters, and how detach prevents gradient leaks.
April 27, 2026
Read article →LangChain Prompt Templates and Output Parsers: PromptTemplate, ChatPromptTemplate, and Pydantic Parsers
Prompt templates and output parsers are the input and output contracts of your LLM pipeline. Build them wrong and your chain breaks on every edge case.
April 27, 2026
Read article →TensorFlow 2.x Architecture: Eager Execution, tf.function, AutoGraph, and Graphs
TensorFlow 2.x made eager execution the default, but tf.function and AutoGraph still power production deployments. Understand when and how graphs take over.
April 20, 2026
Read article →
PyTorch Tensors Deep Dive: dtypes, Device Movement, Memory Layout, and Broadcasting
Tensors are the foundation of every PyTorch model. Master dtypes, device movement, memory layout, and broadcasting rules to eliminate hours of debugging.
April 20, 2026
Read article →LangChain Architecture Overview: Chains, Runnables, LCEL, and the New vs Old API
LangChain's architecture changed fundamentally with v0.1. Learn chains, runnables, and LCEL so you build on the current API — not the deprecated one.
April 20, 2026
Read article →