Benjamin Meisl

Software engineer · founder of Meipath LLC

ben@meipath.com · github.com/TheMizeGuy

13,095 monthly users · bootybaybroker.com
3 platforms in production — web, iOS, desktop
23 public repositories
20+ plugins in the fleet
0 bytes of JavaScript on this page

Studio

I run Meipath LLC, a one-person software studio. I design, build, ship, and operate everything here — web, iOS, desktop — directing a fleet of AI agents built on Claude Code.

Before this: five and a half years of real-time C++ systems and test engineering at General Dynamics, on Linux and VxWorks.

Products

bootybaybroker.com

Ad-supported auction-house price analytics for World of Warcraft, with a companion in-game addon on CurseForge and a desktop app.

Behind it: a scan pipeline holding 2.5 million live listings across 83 realms in both regions, minutes-fresh, in Postgres behind a contract-tested API.

TypeScript · Lua · Tauri · Postgres · Railway

limerino.com

A Twitch chat client for iPhone, built in SwiftUI, live on the App Store since June 2026.

Inline emotes from Twitch, 7TV, BTTV, and FrankerFaceZ (animated, wide, and zero-width), moderator tooling down to the AutoMod queue, Live Activities and a Home Screen widget; VoiceOver, Dynamic Type, and Reduce Motion throughout.

Swift · SwiftUI

berkshirescan.com

Automated police-radio capture and transcription for Berkshire County.

Python · speech-to-text pipeline

Infrastructure

I can run all of this alone because of the agent infrastructure underneath it, built on Claude Code and in production daily.

Policy engine

A Python PreToolUse arbiter that gates every agent dispatch against the model and budget policy table. Agents don't get to disagree with it.

Memory

Four layers: self-hosted RAG across 35 memory spaces, an 800-note Obsidian knowledge vault on 10-minute sync cycles, per-project code memory, and freshness auditing that flags facts gone stale.

The store: 45 GB of Postgres and pgvector, around 2.9 million embedded chunks, including a searchable embedding of the terminal history itself.

Figure 2. The memory stack: the knowledge vault and auto-memory sync into the RAG store every ten minutes; the agent retrieves reranked context from the store; per-project code memory is read in place. agent context RAG store knowledge vault auto-memory code memory Voyage embeddings + reranker Obsidian · 10-min ingest freshness-audited per-project · read in place 10-min ingest retrieve
Fig. 2the memory stack: two layers sync into the store; code memory is read in place.

Conductor–executor orchestration

A conductor model plans and validates; cheaper executor models do the bulk tokens: validation gates, fan-out budgets, git-worktree isolation.

Self-compaction

The agent decides when to compact its own context; a watcher executes the compact and restores state from a session ledger. I only let it do that after watching it earn the call.

Plugin fleet & marketplace

20+ plugins shipped through a self-hosted marketplace with an automated hourly release pipeline and parity-checked public mirrors.

CI/CD fleet

Managed runners plus a self-hosted fleet across two machines, with a placement policy written by a $2k cost postmortem.

Guardrail hooks

Recurring mistakes become machine-enforced rules: a hook layer that warns or denies an agent's tool call before it runs.

Figure 1. A prompt flows to the conductor, which plans; the policy gate checks six invariants before dispatch to N parallel executors; results return through blackboards and a validation gate to the conductor. prompt conductor policy gate executors ×N plan dispatch six invariants blackboards validation gate
Fig. 1one prompt through the orchestration layer.

Open source

The public slice of the infrastructure — plugins, hooks, and skills — is at github.com/TheMizeGuy. Highlights:

Agent teams & review

  • ui-craft ten-specialist UI engineering team for web, iOS, Android, and desktop.
  • apple-ui-craft seven agents for iOS craft: Liquid Glass, spring motion, haptics, accessibility.
  • ios-code-review senior iOS review simulating Apple App Review and a platform engineer, twelve dimensions.
  • typescript-senior-review TypeScript review across eighteen angles with severity-tagged rewrites.
  • typescript-refactor JS-to-TypeScript migration planning: seven read-only planners in waves.
  • api-expert API design, review, and security audit with an embedded knowledge base.
  • cicd-expert pipeline architect, reviewer, debugger, and security auditor.
  • claude-dev autonomous dev team: kanban or ad-hoc, worktree isolation, three-stage review.

Infrastructure & operations

  • railway-operator autonomous Railway operator: CLI-first, seventeen built-in gotchas, verifies every mutation.
  • claude-code-smart-compact the self-compaction hooks: watermark nudges, session ledgers, post-compact rehydration.
  • anti-slop catches and blocks detectable AI patterns in prose, code, and UI.
  • scrum-master kanban manager: story-per-file source of truth, dispatch-wave planning.
  • project-onboard new-project bootstrapper: scaffolder, runbook, verification gate.
  • deep-research multi-agent research with five tiers of collector fan-out.

Domain skills

  • seo-optimization site-agnostic SEO playbook: real authority, zero over-optimization.
  • wow-addon-dev WoW addon development for AI coding agents, retail and Classic.

Writing

Systems write-ups on the infrastructure above; each one ships here as it's finished.

The policy engine in progress

A deny-by-default arbiter between the agents and the tools they call.

Anatomy of a $2k CI bill in progress

A cost-governance postmortem: how six days of managed runners taught the fleet its placement policy.

A memory OS for agents in progress

Why an agent needs four memory layers, and how they stay fresh.

Conductor economics in progress

When a frontier model should plan and a cheaper one should type.

Self-compaction in progress

How the agent earned the right to choose its own compaction points.