Meta Ads + Claude Code: Improve Our AI Media Buyer Learn ($200K/mo Spend)

We've been media buying wtih Meta ads in the insurance space for the past 4 years. We sell leads to insurance agents and currently spend $200,000/month on Meta. Company is closrleads.com Recently we stopped buying manually and built an AI to do it for us. It runs the account on its own: adjusts budgets, turns states on and off, scales winners, cuts losers. It works fine but your job is to use your expertise with media buying to make it better. Two primary areas to optimize 1. Media buying. Campaign management, adjusting budgets, turning states on and off, scaling and cutting. 2. Creative generation. Using performance data to change how creatives get made. We want the system to test new things, learn from what actually works, and improve itself over time without us rewriting its instructions every month. To be qualified for this, you have to be creative. You need real Meta Marketing API experience. You should be fluent with Claude Code. You should have opinions about what makes an autonomous system safe to trust with real money, because this one already has write access to a live account. We asked our AI to explain itself and what it does. That writeup will be shared with serious candidates so you can see exactly what you'd be improving ... but here is the quick rundown -Denis is an agent, not a rules engine. It runs a tool-use loop with about a dozen tools split between reads and writes. -It reads live account data, our own outcome data, and open demand. It stages changes rather than firing them. -Every staged move is reviewed by a separate adversarial model that can kill it before execution. -Underneath that is a hard-coded safety layer enforcing scope, budget caps, and blocklists regardless of what the model decides. -After each run, a second reviewer grades the whole thing against the same data and can trigger one revision round. -There's a schedule, a kill switch, and a full audit log. -Creative generation is a separate pipeline. Performance data feeds the generation prompt. Output goes through human approval. Approved creative gets built paused and needs a separate gated decision before it spends. -Creatives are ranked by what buyers actually closed, not by lead cost. -The infrastructure is solid. The judgment is not fitted. -Denis' decision logic is a long hand-written prompt encoding years of our best buyer's experience as fixed rules and thresholds. When conditions shift, we rewrite the prompt. -Its only persistent memory is free-text notes it writes to itself. -We want those rules fitted, backtested, and self-updating. Same on the creative side, where structured attributes get extracted from past winners but nothing models them against outcomes. Full technical writeup will get sent to serious candidates. We'll also give you access to the dashboard that runs our Meta buying (screenshot below). How this starts Send your initial thoughts and strategy / experience that qualifies you for this. Then, we send you Denis' full explanation of itself so you can see exactly how it works today, and walk through our Meta ads manager over a quick 15 minute call discussing what to change and why. That's really where I can see how you think. If it all lines up, we do a contract and get started.