Home/Case Studies
Technical deep-dives by Sharon Rosario

Case Studies

Security breakdowns, product builds, and engineering notes - written to help you understand the problem and act on it.

Developer Tooling

Antigravity Account Switchers: Why Free Quotas Get Farmed, and What It Costs You

Rotating three or four Google accounts to stack Antigravity free quota is a one-click extension away. I was going to write the how-to. This is what I wrote instead: why the quota exists, how providers fight back, how the switchers work inside, and how to use more than one account properly.

A look at Antigravity account-switcher extensions and free-tier farming: why the quota exists, the same cycle at Cursor and the Gemini API, Google's February 2026 bans and second-strike rule, how profile-swap and proxy switchers work, the security cost of storing every account's login token, and how to use work and personal accounts without the risk.

AntigravityGeminiAI Tools
October 20269 min
Read case study
Developer Tooling

Switch Claude Code Accounts Without Logging Out, and Keep Every Conversation

One command per account, no logout, no OTP, and the same chats, memory and settings on every one of them. Hit a usage limit, switch, and continue the exact same conversation. Step-by-step setup for Windows, macOS and Linux.

How I run several Claude Code accounts on one machine: one shared ~/.claude folder for conversations, memory and settings, and a saved login per account that a small shell function swaps in. Includes the PowerShell, bash and macOS versions, first-time login, and continuing a chat on a different account with --continue.

Claude CodeDeveloper ToolsPowerShell
October 20267 min
Read case study
Developer Tooling

I Built a Claude Code Skill That Tells Me Not to Build Things

An agent will approve every side-project idea you bring it. Here is the open-source skill I wrote to make "don't build this" a possible answer — the gate, the milestone rules, and the one project I ran through all of it.

How idea-to-mvp works: a gitignored profile that makes "worth building" measurable, four kill criteria that can stop a project before any code exists, six context files that keep state on disk instead of in the chat, milestones sized to a single conversation, and a quality loop whose stated success condition is an honest ceiling.

Claude CodeAI AgentsOpen Source
September 202614 min
Read case study
AI Research

I Built a SynthID Watermark Detector That Cannot Detect Gemini

A week reverse-engineering Google SynthID-Text watermarking, a detector that scores ROC-AUC 1.0000, and the one-digit experiment that proved no third party can ever detect Gemini or Claude output.

How SynthID-Text watermarking actually works, what it takes to build a working detector from the Nature paper and Google reference code, and why watermark detection is mathematically impossible without the provider secret key — with benchmark numbers on how editing, paraphrasing and temperature destroy the signal.

SynthIDAI WatermarkingLLM
August 202617 min
Read case study
Security

Understanding & Removing the Git Config Injection Worm

How a self-propagating Git worm spreads across local project folders — detection, removal, and a Windows-first playbook.

Technical breakdown of VS Code auto-task exploitation, PostCSS config injection, and silently rewritten Git commits — with copyable IoCs, audit scripts, and step-by-step removal on Windows.

Malware AnalysisGitWindows
June 202612 min
Read case study
Architecture

Zero-Leak Multi-Tenant RAG Architecture

How to build a B2B AI product that mathematically guarantees one customer can never read another customer's data — using Postgres RLS and pgvector.

A deep technical guide for CTOs and founding engineers on enforcing true data isolation in multi-tenant RAG systems — covering Postgres Row-Level Security, pgvector, namespace partitioning, and a complete security test suite.

AISaaSSecurity
June 202618 min
Read case study
Architecture

Decoupling AI: From Blocking Calls to Event-Driven Microservices

Why synchronous OpenAI API calls will kill your production server under load — and the exact architecture that fixes it using Redis, BullMQ, and WebSocket streaming.

A complete implementation guide for moving from fragile, timeout-prone synchronous AI calls to a scalable event-driven pipeline — with Node.js, Python FastAPI, BullMQ, Redis, and a React streaming frontend.

AINode.jsPython
June 202620 min
Read case study
AI Engineering

Taming Autonomous Agents: Orchestrating Multi-Agent Workflows with LangGraph

How to move beyond unreliable "God prompts" and build deterministic, self-correcting AI agent pipelines that are actually production-ready.

A complete architectural guide and Python implementation for building a Researcher-Writer-Verifier multi-agent pipeline using LangGraph — including state management, self-correction loops, observability with LangSmith, and production hardening patterns.

LangGraphAI AgentsPython
June 202625 min
Read case study