Writing

Blogs

Field notes on AI agents, RAG systems, Claude Code, and the craft of shipping software — long-form deep dives by Giridhar Chettiar.

8Posts

3Topics

92mRead time

Featured

AI & MLJune 13, 202611 min read6

Agent Harnesses — The Model Was Never the Bottleneck

Tejas Kumar took one of the worst models you can still rent — gpt-3.5-turbo — gave it one fixed prompt and one task (upvote a story on Hacker News), and made it reliable without ever touching the prompt or swapping the model. He did it across five git branches by engineering everything around the model: tools, context, guardrails, a verify step, a login handler. This is a hands-on, build-it-with-me walk through that repo and the idea underneath it — an AI harness is everything except the model weights.

AI AgentsAgent HarnessHarness EngineeringContext EngineeringGuardrails

Read post

Browse8 of 8

Agent BattleGround — Claude Code vs Codex

AI & ML

June 11, 202618 min read9

Agent BattleGround — Claude Code vs Codex

I gave Claude Code (Fable 5) and Codex (gpt-5.5) the exact same six creative-coding tasks — an arcade game, a generative-art CLI, a bug hunt, an ASCII aquarium, a messy-data dashboard, and a 120-line demoscene — one shot each, zero interventions, blind judges, instrumented re-tests. This is the full fight card: every prompt, the actual screen recordings, the parameter-by-parameter scorecards, and what the judges said. Final tally: 6–0, but the story is closer than the headline.

Claude CodeCodexFable 5+6

AI & ML

June 6, 202614 min read5

MCP — The What

Part 2 of a 3-part series on MCP. Part 1 was the why; this is the anatomy. We open up a single connection between an AI and a tool and name every part — Host, Client, Server, the three primitives a server offers, the exact JSON-RPC that crosses the wire, and the transport (STDIO vs HTTP+SSE) it crosses through. Four interactive diagrams you can actually drive.

MCPModel Context ProtocolArchitecture+7

AI & ML

May 31, 202611 min read3

MCP — The Why

Part 1 of a 3-part series building your intuition on MCP. Why connecting AI to your tools quietly turns into an integration explosion — N tools times M services — and how the Model Context Protocol collapses that to N + M. Includes an interactive visualisation and a preview of the MCP server I built to solve my own context problem.

MCPModel Context ProtocolAnthropic+4

AI & ML

May 28, 20268 min read2

Skills in Claude Code

Why a general-purpose model still struggles on specific tasks, what a Skill actually is, and how progressive disclosure makes a library of them practical — written up from My Personal Project Experience, the CampusX video and Anthropic's own docs.

Claude CodeSkillsAnthropic+2

Performance Testing

May 16, 20258 min read1

On Automating Performance Testing with JMeter: A Headless CLI Approach

Learn how to set up a fully automated JMeter testing environment for continuous performance monitoring with minimal overhead and maximum efficiency.

JMeterPerformance TestingAutomation+2

Configuring a Next.js website with Sanity CMS with blogs and Dynamic Page rendering

Web Development

Apr 15, 202412 min read

Configuring a Next.js website with Sanity CMS with blogs and Dynamic Page rendering

Learn how to integrate Sanity CMS with Next.js for dynamic content management, blog creation, and flexible page rendering

Next.jsSanity CMSContent Management+1

Exploring Gen AI the right way without the hype

AI & ML

Apr 18, 202310 min read2

Exploring Gen AI the right way without the hype

Navigate the world of Generative AI beyond the hype, focusing on practical applications and responsible implementation

Generative AIAI EthicsMachine Learning+1

Want to talk shop?

I write about what I'm building and breaking. If something here sparked an idea, let's trade notes.

Get in touch