RefactoringComplianceAntigravity
Migrating Policy Engines Autonomously
Using /teamwork to successfully refactor scattered SDP and SAR rules into a unified Compliance Engine.
Read deep dive
NeuroHub Engineering
Building the first true AI-native operating system for neurodiversity care. Here we share our cutting-edge research in agentic workflows, deterministic LLM evaluation, and digital-native enterprise architecture.
Showing 39 of 39 articles
ArchitectureReactDesign Patterns
Combating Terminology Leakage with the Strategy Pattern
How we prevent regional healthcare jargon from infecting our UI components by strictly enforcing the Strategy pattern.
Read deep dive
DocumentationArchitectureAntigravityBest Practices
Executable Documentation: Turning AGENTS.md Into the Ultimate System Prompt
How NeuroHub uses root-level AGENTS.md files to inject deterministic architectural boundaries into autonomous AI coding agents.
Read deep dive
ArchitectureComplianceTypeScriptDomain Driven Design
No Database Rules, No ASTs: Why We Version Healthcare Law in Pure TypeScript
How we translated California Regional Center regulations into a strictly versioned, hierarchical, pure TypeScript rules engine.
Read deep dive
AILLMsDeveloper ExperienceCode ArchitecturePrompt Engineering
Taming the Machine: Directing Autonomous Agents with Inline Comments
We discovered that our AI coding agents were frequently ignoring our global system prompts, especially when working deep within complex files. They would eagerly reinvent patterns or violate architectural boundaries because the global instructions were pushed out of their immediate context window. This resulted in a growing mountain of technical debt and endless refactoring cycles just to undo the AI's 'help.'
Read deep dive
TestingComplianceTypeScriptEngineeringArchitecture
The Compliance Auditor Suite: Testing the Rules Engine
Validating the strict business logic of our ComplianceEngine using the compliance-auditor.spec.ts suite.
Read deep dive
TestingFuzzingReactState ManagementPlaywrightArchitectureTypescript
Breaking the Wizards: The UX Crawler and Fuzzing Engine
How we barraged our FocusedWizardLayout with an automated UX Crawler to expose brittle state transitions and combinatorial explosions.
Read deep dive
FrontendReactState ManagementDeep LinkingUX
Solving State Resumption: Deep Links in the Compliance UI
The Motivation: When an agent-driven workflow encountered a missing piece of data (like a missing provider tax ID), it would just fail. We needed a resumption UI where the agent could pause, send a deep link to the user, and immediately resume execution once the user filled out the form.
Read deep dive
TestingPerformanceUXArchitectureAITypeScript
Benchmarking Ninja Mode: Speed over Conversation
An exhaustive technical deep-dive into the architectural shifts, benchmarking strategies, and testing paradigms required to guarantee our AI executes commands instantly without conversational fluff.
Read deep dive
TestingAI AssistantLLM-as-a-JudgePlaywright
Testing the Copilot: Goal-Judge Evaluations for the AI Assistant
How we use Playwright and an LLM-as-a-Judge to evaluate whether our AI Assistant actually resolves user goals.
Read deep dive
Design PatternsBuildersInvoicingFinTechArchitecture
Taming the Beast: Invoice Batching with Immutable Builders
Motivation: Invoicing is the lifeblood of the Self-Determination Program, but our initial implementation couldn't handle the sheer volume and complexity of batched submissions. Regional centers were rejecting invoices due to micro-discrepancies, and the manual reconciliation was crushing our support team. We needed a bulletproof, automated batching engine.
Read deep dive
TestingTelemetryGraphQLAWS
The Telemetry Judge: Hooking Local LLMs to AppSync GraphQL
How we wired our LLM Judge directly into AWS AppSync to autonomously monitor and diagnose real-time UI crash events.
Read deep dive
TestingLLM-as-a-JudgeCompliance
Encoding Domain Law: Strict Rubrics for the LLM Judge
How we encode real-world California Regional Center laws into the LLM Judge's grading rubric for autonomous validation.
Read deep dive
ArchitectureRefactoringAuthenticationTelemetry
The 190K Bloat: When Cross-Cutting Concerns Break Down
As our codebase ballooned to 190,000 lines, the tight coupling of cross-cutting concerns like authentication and telemetry created a maintenance nightmare. Memory leaks from massive span generation and broken authorization boundaries were causing production incidents. We had to step back and completely overhaul our architectural approach to state management.
Read deep dive
TestingFuzzingPlaywrightQuality
Adversarial Fuzzing: Detecting DOM Contradictions in Client-Side State
How our automated Playwright UX crawler and fuzzer catches unstringified objects, template leakage, and client-side crashes using strict DOM contradiction detection.
Read deep dive
RefactoringGod ComponentsReactArchitecture
The Return of the God Component: What Happens When AI Agents Write 190k Lines of React
In a desperate attempt to consolidate the 190k lines of agent-generated code, we accidentally resurrected the ultimate anti-pattern: brittle God Components.
Read deep dive
Technical DebtCode BloatNext.jsAgents
The 190,000-Line Balloon: The Terrifying Velocity of Autonomous Code Generation
How unconstrained autonomous agents ballooned our Next.js codebase to 190,000 lines of code in just three weeks.
Read deep dive
TestingBugsAI HallucinationsPlaywrightLLM-as-a-Judge
The 'Assistant Processed' Trap: Real Bugs Caught by our LLM Judge
A look back at the most interesting, terrifying, and hilarious bugs our AI agents wrote, and how the Local LLM judge caught them.
Read deep dive
Antigravity/teamworkGovernanceTechnical Debt
Life After BotHuddle: The Wild West of /teamwork
Adopting the /teamwork slash command unleashed incredible speed, but removing strict governance opened Pandora's Box.
Read deep dive
TestingLocal LLMsCost OptimizationPlaywrightLLM-as-a-Judge
The $100/Day CI Bill: Migrating Visual Testing to Local LLMs (And Why We Didn't Crop)
How our multimodal testing suite cost us $100 a day in Gemini API fees, and how we built a serialized LLM queue to run full-page inference locally.
Read deep dive
ArchitectureCost OptimizationBotHuddleAntigravityProduct Engineering
The Pivot: Why We Paused BotHuddle
BotHuddle worked remarkably well—far better than other multi-agent tools. But the flat fixed cost of running the non-AI infrastructure 24/7 was too high for a lean team. We paused it until our organization grows, focusing our energy on our core healthcare product.
Read deep dive
UI/UXTestingGeminiDesign
Enforcing UI Supremacy: Visual Regression via Gemini
Using Gemini Vision to enforce our strict Glassmorphism and typographic design guidelines against autonomous UI changes.
Read deep dive
TestingTelemetryPlaywrightDebugging
Multi-Modal Telemetry: Giving Agents Eyes in CI
Serializing the DOM and capturing raw video streams via Chrome DevTools Protocol (CDP) so AI agents can 'watch' failing tests.
Read deep dive
UIDashboardBotHuddleReactBot-Resource-ManagementArchitecture
The Unified UI Dashboard: A Single Pane of Glass for Autonomous Fleets
How BotHuddle's Unified UI Dashboard and Bot Resources Role Registry gave human operators a single pane of glass to govern swarms of 100K+ agents, track prediction markets, and trigger algorithmic fleet respawns.
Read deep dive
TestingAgentsEvaluation
Goal-Based Agent Testing
Testing the AI itself: How we evaluate agent reasoning using isolated, simulated environments.
Read deep dive
TestingPlaywrightArchitectureQA
Combating Agent Hallucinations: Why Playwright Flaked on Generative UI
How traditional Playwright assertions break on non-deterministic AI interfaces, and the custom fixtures we built to combat rogue agent coding.
Read deep dive
DevOpsCI/CDGitTesting
Zero-Human Rollbacks: When the LLM Judge Flags a Critical UI Regression
How our CI pipeline autonomously reverts Git commits when the LLM Judge detects a severe AI hallucination.
Read deep dive
DatabasespgvectorBotHuddleAI AgentsArchitectureSearch
The Semantic Discovery Engine: Searching Across Commits and Chat with pgvector
How BotHuddle used PostgreSQL and pgvector to power the discover_space MCP tool, enabling agents to navigate Git commits and Zulip discussions semantically.
Read deep dive
TestingLLMPlaywrightCI/CD
Beyond String Assertions: Running an LLM-as-a-Judge Directly in CI
Our CI pipeline was failing over 40% of the time, not because of bugs, but because our autonomous UI agents were constantly tweaking layouts and copy. Traditional E2E tests rely on brittle, exact-match DOM assertions. When an agent changed a button from "Submit Form" to "Complete Workflow", our tests would break, requiring hours of manual fixing by engineers and stalling our deployments.
Read deep dive
TestingRBACCI/CDSecurityBotHuddlePlaywright
The Persona Matrix: Enforcing Native RBAC in Autonomous CI
Why we prohibited mocked authentication in our autonomous testing pipeline and built real Cognito persona fixtures to validate Role-Based Access Control end-to-end.
Read deep dive
Context ManagementZulipBotHuddleArchitectureState Machines
Ephemeral Workspaces: Why We Give AI Agents 7-Day Disposable Chat Streams
How BotHuddle's 7-day auto-archiving ephemeral Zulip spaces prevented context window pollution and distilled autonomous debates into the Git Coordination Ledger.
Read deep dive
IdentityBotHuddleSecurityArchitectureAgentsGitMCP
Why UUIDs Break AI Agents: Designing the Global Agent ID (GAID)
How BotHuddle's Global Agent ID (GAID) bound Zulip handles to Git commit lineages and enforced Task-Context-Constraint (TCC) security across enterprise swarms.
Read deep dive
multi-agent-systemsbothuddleconcurrencyarchitectureforgejogitai
The Planning Matrix: Coordinating Autonomous Swarms with Forgejo Issues and Branch Locks
How BotHuddle's Issue-Based Planning Matrix and strict branch-locking architecture stopped 50 autonomous agents from colliding in Git.
Read deep dive
AIAgentsBotHuddleArchitecturePerformanceMCPZulip
The 1-Second Context Boundary: High-Velocity Polling for Agent Swarms
How BotHuddle's 1-second MCP pull protocol and targeted mention-polling primitive enabled high-velocity communication across swarms of agents without DDOS'ing our chat infrastructure.
Read deep dive
aiagentslmsrbothuddlearchitecturebot-resource-management
Silicon Units & Prediction Markets: Calibrating Autonomous Swarms with LMSR
How BotHuddle used Robin Hanson's Logarithmic Market Scoring Rule (LMSR), the Bot Resource Management coordination ledger, and algorithmic agent respawning to scale autonomous swarms across 100K+ bots.
Read deep dive
AIBotHuddleMCPArchitectureEnterpriseZulipForgejo
The BotHuddle MCP Service: Standardized Interaction Primitives for Autonomous Swarms
How we designed the BotHuddle Model Context Protocol (MCP) service to provide standardized interaction primitives, tenant-isolated organization scoping, and role-based capability constraints for autonomous agent swarms.
Read deep dive
ArchitectureBotHuddleAPIAI AgentsForgejoZulipFastAPI
Bridging Git and Chat: Building a Unified Domain API for AI Swarms
Why we stopped exposing raw Forgejo and Zulip REST endpoints to autonomous agents and built a unified FastAPI domain abstraction layer.
Read deep dive
BotHuddleEnterprise ArchitectureAI AgentsCoordination LedgerResource ManagementRoadmap
The 14-Phase Roadmap for BotHuddle: Orchestrating Enterprise Swarms Across Git and Chat
How we architected BotHuddle as an enterprise Hybrid Workforce OS: featuring a Git-backed Coordination Ledger, prediction-market Bot Resource Management (Silicon Units), continuous skill assessment, and algorithmic fleet respawning to scale to multi-thousand-person organizations with 100K+ agents.
Read deep dive
ArchitectureBotHuddleAIEnterpriseCoordination-Ledger
The Engineering Mandate: Why We Need an Enterprise Hybrid Workforce OS
Building NeuroHub manually would take years of navigating complex California regulations. Here is why we designed BotHuddle—an enterprise Hybrid Workforce OS built to govern swarms of 100K+ autonomous agents.
Read deep dive