image
Alpesh Kumar

Blog

The UUID Primary Key Problem That Shows Up at 200 Million Rows — a framed technical comparison shows a random UUID B-tree scattering inserts and a point lookup fanning into many cold row pages, contrasted with a time-ordered UUID path that keeps inserts and fetched pages locally clustered. Visual direction: idea: a UUID primary-key lookup can find the key quickly yet still cause extreme page reads when random insertion order scatters the B-tree and cold row pages; subject: UUID primary keys at 200 million rows; mechanism: compare random UUID insert distribution across many B-tree leaf pages with time-ordered UUID inserts concentrated at the active edge, then show a point lookup descending the index and fanning out into scattered cold data pages marked by a dense 94,000-read texture; composition: framed comparison — a generous title band across the top, with a large left diagnostic zone for random UUID scatter and a right diagnostic zone for ordered UUID locality, joined only at the bottom by one shared lookup-to-row path; palette: deep forest/moss — pale eucalyptus background, forest-green and near-black ink, muted sandstone page blocks, restrained burnt-orange I/O accents, and a small burgundy warning accent; exclusions: dark navy, neon-blue glow, floating UI cards, network nodes, generic AI circuitry, people, robots, hands, dashboards, logos, fake code, secondary readable text, decorative charts, physical-machine metaphors

The UUID Primary Key Problem That Shows Up at 200 Million Rows

Alpesh Kumar4 August 2026,(0)
A primary-key lookup can still become expensive when random UUIDs, cache pressure, and table layout collide. Here is how to diagnose the evidence...
Read More
AI Agent Frameworks Compared: Vercel AI SDK, LangGraph, CrewAI, AutoGen and LangChain — a framed comparison diagram visually separates streaming provider-flexible TypeScript output, branching and resumable orchestration, role-based collaboration, conversational agents, and modular integration building blocks, showing why framework selection follows architecture rather than a universal winner.

AI Agent Frameworks Compared: Vercel AI SDK, LangGraph, CrewAI, AutoGen and LangChain

Alpesh Kumar14 August 2026,,,(0)
A practical, source-led comparison of five AI frameworks. Choose from your application architecture, streaming needs, and agent workflow rather than a...
Read More
AI Watermarks Are Not a Silver Bullet: What Provenance Signals Can and Cannot Prove — a warm ivory technical diagram shows a media file with an embedded watermark band and a linked provenance record feeding an evidence tray, then stopping before separate source, context, and corroboration checks, representing that C2PA metadata and watermarks inform origin but cannot independently prove accuracy, ownership, or context. Visual direction: idea: provenance signals are useful evidence but stop short of editorial truth; subject: C2PA Content Credentials and embedded AI watermarking; mechanism: a media file carries an embedded watermark while a separate provenance record supplies origin/history information, both entering an evidence tray before a hard boundary requiring source, context, and corroborating-evidence checks; composition: landscape editorial explainer card with the exact headline in the upper third and a lower horizontal proof-boundary schematic; palette: warm ivory, ink charcoal, muted brick red, moss green, soft ochre, pale stone gray; exclusions: dark navy, neon-blue glow, floating UI cards, network nodes, generic AI circuitry, people, robots, hands, dashboards, physical-machine metaphors, decorative charts, logos, secondary readable text.

AI Watermarks Are Not a Silver Bullet: What Provenance Signals Can and Cannot Prove

Alpesh Kumar13 August 2026,(0)
C2PA metadata and embedded signals can offer helpful evidence about an image's origin. They cannot, on their own, prove that an image is accurate...
Read More
Your First Personal AI Agent: A Focused Weekend Build — a warm cream technical schematic shows current-week meeting notes constrained by an agent contract, optionally checked against one read-only source, transformed into a structured priorities draft, and stopped at a human approval boundary before any external action; visual direction: idea: a first personal AI agent produces a reviewable weekly-priorities draft from tightly bounded meeting notes, with no autonomous external action; subject: focused personal AI-agent weekend build; mechanism: current-week notes pass through a clearly labelled contract boundary and one limited read-only lookup into a structured priorities draft, then stop at an explicit human approval gate before any external-action area; composition: landscape editorial explainer card with the exact title large across the upper third and a lower left-to-right paper-like workflow schematic, ending at a prominent approval stamp boundary and a muted crossed-out external-action zone; palette: warm cream background, charcoal typography, oxblood red boundary accents, muted olive green for approved draft, dusty apricot for source notes, slate gray connectors; exclusions: dark navy, neon-blue glow, floating UI cards, network nodes, generic AI circuitry, people, robots, hands, dashboards, physical-machine metaphors, decorative charts, logos, secondary readable text.

Your First Personal AI Agent: A Focused Weekend Build

Alpesh Kumar2 August 2026,,,(0)
Build a small, reviewable AI workflow in a couple of hours by giving it one narrow job, clear boundaries, and a human approval step...
Read More
Build an AI Agent Evaluation Flywheel That Improves Prompts — a text-free circular technical schematic shows recorded agent traces being separately evaluated for outcomes and trajectories, classified into recurring failure types, turned into prompt candidates, and checked against development, hold-out, and regression cases before feeding evidence back into the next run. Visual direction: idea: a trace-aware evaluation loop turns recurring agent failures into verified prompt changes; subject: AI agent evaluation flywheel; mechanism: captured tool traces flow through separate outcome and trajectory evaluation, failure taxonomy, prompt variants, and hold-out/regression verification with a safety gate; composition: a single clockwise five-stage circular schematic with evidence records at the top and a verification boundary closing the loop; palette: warm off-white, graphite, deep forest green, muted rust, clay orange, restrained golden yellow; exclusions: dark navy, neon-blue glow, floating UI cards, network nodes, generic AI circuitry, people, robots, logos, readable artwork text.

Build an AI Agent Evaluation Flywheel That Improves Prompts

Alpesh Kumar1 August 2026,,(0)
A five-stage evaluation flywheel for turning AI agent traces into evidence-backed prompt improvements, without treating final answers as the whole...
Read More
Why Final-Answer Evals Leave AI Agent Failures Invisible — a warm editorial execution-trace schematic contrasts a reassuring green final-response indicator with the visible failed path beneath it: wrong tool selection, a quoted-number parameter mismatch, and a missing update state, all examined by a trajectory evaluator. Visual direction: idea: an apparently successful answer hides a failed execution path; subject: AI-agent trajectory evaluation; mechanism: evaluator compares final claim with recorded tool selection, typed arguments, missing action, and state transitions; composition: left-to-right audit strip beneath a detached green outcome capsule with an inspection bracket; palette: warm ivory, charcoal, muted sage, terracotta, ochre; exclusions: dark navy, neon-blue glow, floating UI cards, network nodes, generic AI circuitry, people, robots, logos, readable artwork text.

Why Final-Answer Evals Leave AI Agent Failures Invisible

Alpesh Kumar31 July 2026,,,,(0)
A correct final response can hide failed tool calls, broken constraints, and bad state changes. Learn why AI agent evaluation must score the full...
Read More
Vector Quantisation: How Code Books Compress Similar Data — an illustrated flow of many similar input vectors being matched to a small numbered code book, then represented by one compact index to show lossy shared-dictionary compression.

Vector Quantisation: How Code Books Compress Similar Data

Alpesh Kumar30 July 2026(0)
Vector quantisation replaces each group of values with the index of a representative vector. Learn how code books, k-means, distortion, and trade-offs...
Read More
TurboVec: What This Rust Vector Index Is and How to Evaluate It — a dark technical illustration shows a glowing field of connected vector points compressing into an organized geometric index as a bright query path selects nearby vectors, representing evaluation of compressed vector retrieval.

TurboVec: What This Rust Vector Index Is and How to Evaluate It

Alpesh Kumar26 July 2026,,(0)
TurboVec is a Rust-based vector index with Python bindings built on TurboQuant. This article explains what the project claims to be, where vector...
Read More
Context Engineering for AI Agents: Memory, Retrieval, and Token Budgets — an abstract AI-agent core receives separately filtered memory, retrieved evidence, and segmented token-budget streams, representing deliberate context selection for the next decision.

Context Engineering for AI Agents: Memory, Retrieval, and Token Budgets

Alpesh Kumar23 July 2026,,,(0)
A practical framework for developers building AI agents that need the right context at the right time—without drowning the model in stale history...
Read More
Clean Architecture: Where SOLID, DDD and Event-Driven Systems Fit — a layered architectural illustration shows a protected business-rule core, modular domain rings, outward event particles, and separate database, API and messaging infrastructure to represent inward dependencies and decoupled follow-up work.

Clean Architecture: Where SOLID, DDD and Event-Driven Systems Fit

Alpesh Kumar24 July 2026,(0)
Clean Architecture is not a framework or a checklist. This practical guide shows how SOLID, domain-driven design and events can work together—without...
Read More
Multi-Agent Systems: Your Guided Learning Path — a visible connected route through distinct AI-workflow stations, guardrails, and a human-reviewed destination represents the article’s structured learning path from foundations to responsible decisions.

Multi-Agent Systems: Your Guided Learning Path

Alpesh Kumar25 July 2026,,,(0)
A beginner-friendly roadmap to understanding multi-agent systems, from the essential concepts to coordination, guardrails, and deciding whether to...
Read More
What Are Ensemble Methods in Machine Learning? A luminous central prediction node receives distinct streams from decision trees, sequential learners, and model modules, showing diverse models combining into one dependable result.

What Are Ensemble Methods in Machine Learning?

Alpesh Kumar1 February 2022(0)
Ensemble methods combine predictions from multiple machine-learning models to produce a more reliable result. Learn how bagging, boosting, stacking...
Read More
Alpesh Kumar
Powered by  GDPR Cookie Compliance
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.