All posts by Alpesh Kumar
Your First Personal AI Agent: A Focused Weekend Build
Build a small, reviewable AI workflow in a couple of hours by giving it one narrow job, clear boundaries, and a human approval step.
Build an AI Agent Evaluation Flywheel That Improves Prompts
A five-stage evaluation flywheel for turning AI agent traces into evidence-backed prompt improvements, without treating final answers as the whole story.
Why Final-Answer Evals Leave AI Agent Failures Invisible
A correct final response can hide failed tool calls, broken constraints, and bad state changes. Learn why AI agent evaluation must score the full trajectory.
Vector Quantisation: How Code Books Compress Similar Data
Vector quantisation replaces each group of values with the index of a representative vector. Learn how code books, k-means, distortion, and trade-offs fit together.
Contextual Retrieval: Anthropic’s Approach to Reducing RAG Retrieval Failures
Anthropic’s Contextual Retrieval adds document-level meaning to RAG chunks before embedding and BM25 indexing. Here is how it works, what the 67% result means, and when to use it.
TurboVec: What This Rust Vector Index Is and How to Evaluate It
TurboVec is a Rust-based vector index with Python bindings built on TurboQuant. This article explains what the project claims to be, where vector compression fits, and what developers should verify before adopting it.
Multi-Agent Systems: Your Guided Learning Path
A beginner-friendly roadmap to understanding multi-agent systems, from the essential concepts to coordination, guardrails, and deciding whether to build one.
Clean Architecture: Where SOLID, DDD and Event-Driven Systems Fit
Clean Architecture is not a framework or a checklist. This practical guide shows how SOLID, domain-driven design and events can work together—without turning a straightforward application into an over-engineered maze.
Context Engineering for AI Agents: Memory, Retrieval, and Token Budgets
A practical framework for developers building AI agents that need the right context at the right time—without drowning the model in stale history, irrelevant documents, or runaway token costs.