AI engineering consultant to audit how I build software with AI agents, top to bottom

I run a small service business and I am not a professional engineer. Over the past few months I have built our operating software, a web and desktop app the team uses every day, almost entirely with AI coding agents. It works and we depend on it, but I suspect I am paying for a lot of waste. I want an experienced outsider to look at the whole operation and tell me how to do this better. How it runs today - Claude Code, with up to six parallel builders working from a ranked queue in separate git worktrees - One directing thread (my "front door") that I talk to, which routes work to the builders - A mix of Sonnet, Opus and Fable models, on flat monthly plans of about $600 - Roughly 150 custom hook files, about 240 written decision records, and a memory system - Next.js, React and Electron, with Python and FastAPI on Postgres, hosted on Railway, with GitHub Actions for CI Why I want this - About 1.7M lines of code, up from about 565k in August - 450 to 1,000 merged PRs a week, nearly all written by agents - Over the last three weeks, the CI success rate fell from 77% to 45% and median CI time rose from about 11 to about 18 minutes - Fixes often seem to cause new problems somewhere else, and I can't tell how much of the volume is real progress What I want your judgment on 1. Models: which model and effort level should do which work? Am I paying for quality I don't get? 2. Cloud versus local: what should run on local models or my own hardware, and what should stay in the cloud? This covers both AI work and CI. I have a 128 GB Mac on order. 3. Directing process: is the front door and queue setup a good way to run this, or is it overhead? Are six parallel builders too many or too few? 4. Guardrails: which hooks, rules and memory earn their cost, and which should go? 5. Codebase size: is 1.7M lines a problem in itself, and what should be done about it? 6. Quality and speed: why is CI degrading, and why do bugs keep returning? 7. My own habits: what am I doing as the operator that makes all of this worse or more expensive? What you deliver - A written report, ranked by impact, with specific evidence from my setup - What to stop, keep and change first, with estimated savings or speedup for each - A 60-minute call to walk me through it - Your read on whether I need a human engineer at all, and what kind I am paying for your independent judgment, so tell me plainly if I am doing this wrong. Please do not sell me a framework, a platform or an ongoing retainer as part of this work. Access and confidentiality Read-only access to the code, CI history and agent configuration, plus usage data, under an NDA. No production credentials and no customer data. You must have - Run multi-agent coding workflows yourself on a real production codebase. Chatbots and prompt writing don't count. - Hands-on experience with Claude Code or a similar tool, including hooks or equivalent controls - Enough engineering background to read CI data and a large codebase critically - Real experience comparing AI model cost against output quality To apply, tell me - The largest agent-built system you have personally run or audited, with its rough size - A time you told a client to use less AI or spend less on it - Which of the seven questions you would look at first, and why - Your fixed price, timeline and availability Generic proposals will be skipped. Please start your reply with the word "worktree." Budget and timeline: I have not set a budget because I don't know what this should cost. Please quote your price and estimated hours based on what you read here. I would like this finished within one to two weeks.