Blog

Technical writing on backend systems, architecture decisions, and engineering trade-offs.

8 min read

The Agentic Flip: Why I Ditched Heavy Frontier Models for Gemini 3.5 Flash and Haiku

High-volume agentic coding doesn't require multi-dollar frontier models. By combining a strict local orchestrator (Plan, Implement, Test, Review) with ultra-fast, cheap models like Gemini 3.5 Flash (Medium) and Claude 4.5 Haiku, I cut AI costs while boosting iteration speed.

10 min read

Building a Serverless Comment System: Astro, Giscus, and GitHub Actions Automation

A deep dive into integrating the Giscus comment system into Astro 4.x, resolving View Transitions lifecycle bugs, and automating GitHub Discussions pre-creation via the GraphQL API.

6 min read

Microsoft MAI Debuts: The Dawn of Redmond's AI Independence

At Build 2026, Microsoft launched the MAI model family—marking a major pivot toward in-house model ownership and 'humanist superintelligence.' Here is a technical teardown of the models and the strategic shift.

7 min read

6,000 Credits in an 8-Hour Workday: The Broken Economics of Copilot Enterprise

How a single day of agentic coding evaporated a corporate AI budget. An analysis of GitHub's new metered billing, why paying 1:1 API rates inside a premium SaaS license is unsustainable, and what it means for engineering organizations.

8 min read

LLM Headroom: Slashing Token Consumption by 60–95% in Agentic Workflows

AI agents running inside tools and loops are incredibly resource-intensive. Learn how chopratejas/headroom uses reversible compression and AST-aware tools to optimize LLM contexts and slash token costs.