3 ms·

Show HN: Accidentally built churn detection AI (I don't code)

2 points by willster787 1y ago

I'm new here, so apologies in advance for link faux pas, etc. Six months ago, I discovered Claude could turn my ideas into code. It started with batch files at work - I needed automation, couldn't find scripts, asked ChatGPT to fix snippets. Ran out of tokens, switched to Claude. It got things right the first time. My boss was impressed. Then I thought: "Could I build something to sell online?" Started building for realtors (they have money, need tools). Got deep into ML predicting housing markets. Asked Claude "can this AI make AI?" and... it could. My realtor brother-in-law killed it: "Market's too chaotic, this won't help." Pivoted to his simpler idea but kept the AI parts. Built an "AI assistant" that was supposed to be a small feature. It became the whole system. 80% done, tried to test it. Wouldn't even start. Spent WEEKS copy-pasting errors to Claude, entire 2000+ line files, racing against token limits. Finally powered on. New errors meant progress. Then reality hit: I'd built an advanced decision-making system that needed months of data from hundreds of agents. They just wanted a chatbox. Asked Claude: "What is this actually good for?" Answer: "Churn prevention. Multi-billion dollar problem." Three weeks ago, discovered Claude Code (MCP). Plugged it into my folder. Game changer. Now I have: - Real-time churn detection (85% accuracy on "I want to cancel") - Claude integration that's aware of the "temperature" of each conversation - Self-healing architecture (because I can't fix what breaks) - SHAP explanations showing why predictions were made - 10,000+ lines of production TypeScript The kicker: I still don't know how to code. I just kept asking "what if?" curl -X POST http://localhost:4000/api/chat/message \ -d '{"message": "I want to cancel", "userId": "test-user-001"}' Response in 1.8 seconds with full risk analysis. Looking for one SaaS company to actually use this. I built something that predicts human behavior eerily well. Turns out predicting when customers leave is worth real money. Who knew? Currently running locally - happy to demo for interested companies or give temporary API access to serious testers. Didn't want to wait until I had my DNS act together before sharing this. UPDATE: Live demo available! https://donnabrain.com/chat https://donnabrain.com/chat Type "I want to cancel my subscription" and watch it detect 85% churn risk with real-time explanations. You can try other prompts as well, obviously. Running on an RTX 3090 in my home office. The system you're hitting is the actual production code from the README. No signup required - just hit "Start New Session" on top and start typing.

vintagedave 1y ago
Your story is great, but may make people think of the product as unreliable - it doesn't sound ready to be a SaaS yet. I tried it out, sending two messages, saying I hadn't had service for 2 weeks and then saying I wanted someone to come look at it. It was a steady 50.1% churn risk. Not sure that's high enough. That said: I think it has real potential. Polish the demo up (eg even having to click 'start new session' felt very tech-ish, not demo-ish, if you know what I mean? Reduce the tech load and increase the simple smooth dashboard feel), see if you can make it a deployable system easy for an admin, or could spin up your own servers and customer accounts. Good luck!
- willster787 1y ago
  First of all- thank you so much for giving it a try. I know it's a bit rough around the edges. I found what I think is some really solid data and I'm going to try to train it tonight. The "idea" was to show what it could possibly do for a SaaS Dashboard, I was thinking "clippy but not useless". REALLY appreciate the constructive criticism!

2 points by willster787 1y ago

vintagedave 1y ago

Your story is great, but may make people think of the product as unreliable - it doesn't sound ready to be a SaaS yet. I tried it out, sending two messages, saying I hadn't had service for 2 weeks and then saying I wanted someone to come look at it. It was a steady 50.1% churn risk. Not sure that's high enough. That said: I think it has real potential. Polish the demo up (eg even having to click 'start new session' felt very tech-ish, not demo-ish, if you know what I mean? Reduce the tech load and increase the simple smooth dashboard feel), see if you can make it a deployable system easy for an admin, or could spin up your own servers and customer accounts. Good luck!

willster787 1y ago
First of all- thank you so much for giving it a try. I know it's a bit rough around the edges. I found what I think is some really solid data and I'm going to try to train it tonight. The "idea" was to show what it could possibly do for a SaaS Dashboard, I was thinking "clippy but not useless". REALLY appreciate the constructive criticism!