From inside MaidLink
Two series published on LinkedIn. Field Notes tracks what we are building and why. PM Signal covers how I think about AI product decisions. Written in the middle of building, not after.
Enterprises stopped asking 'Can we use AI?' They started asking 'Can we use AI at scale?' That shift changes everything about what AI PM work actually is. Read on LinkedIn →
A fully remote company just hit $300M ARR. No offices. No in-person. Just a product people needed and a team that shipped. The model works — if the product is real. Read on LinkedIn →
Nothing we tried worked. More parameters. Different questions. Adjusted the logic. Ran it again. The estimator kept coming back wrong. Read on LinkedIn →
AI startups are raising on the story. Clients are buying on the demo. And somewhere between the pitch deck and production, the actual product has to work. Most don't. Read on LinkedIn →
Most AI products ship without a definition of 'working.' We had to write ours down first. MaidLink's estimator reads home condition from photos. We replaced client self-input entirely. Read on LinkedIn →
Google's AI agents are starting to handle SMB queries directly. The referral traffic that kept small marketplaces alive is shrinking. That changes who needs to own discoverability. Read on LinkedIn →
An engineer can give you the accuracy number. Only the PM can decide if it is good enough to ship. Read on LinkedIn →
We thought we had training data. We had operational data. They are not the same thing. 150 jobs. Time logs. Photos. Enough to run a business. Not enough to teach a model. Read on LinkedIn →
Most AI features ship and quietly decay. The model was trained on yesterday's data. The world moves. The model doesn't. Nobody notices until the estimates are consistently wrong. Read on LinkedIn →
Match Group just posted for AI PMs across every brand. Not engineers to build AI. PMs to own it. That shift — from AI as an engineering project to AI as a product discipline — is accelerating. Read on LinkedIn →
Most teams pick the most accurate model and ship. We evaluated three — Nova, Claude, and GPT-4o — for MaidLink's estimator. Claude was most accurate. We chose Nova. Read on LinkedIn →
PayPal announced $1.5B in savings from AI. The headline is the number. The story is how they got there: narrow use cases, high-confidence outputs, clear fallback paths. That's not magic. That's product discipline. Read on LinkedIn →
The app hit its download record last week. Bookings were flat. That gap — installs without conversion — is a matching problem, not a marketing problem. The supply side wasn't ready for the demand. Read on LinkedIn →
The plan was wrong. Not the vision. The plan. We had a marketplace built over 18 months. Real clients, real cleaners, real bookings. And a process that kept breaking in the same place. Read on LinkedIn →
AWS Bedrock just added four new models. Every time a platform adds models, the same question resurfaces: do you build model-agnostic, or do you go deep on one? We went deep on one. Here's why that was the right call for where we are. Read on LinkedIn →
We built the marketplace. Then we stopped. Not because the idea was wrong. Because after 150+ real jobs, we could see exactly what wasn't ready. Read on LinkedIn →
Nobody tells you what it actually feels like to build something with no fallback. Not the inspirational version. The real one. Read on LinkedIn →
A couple of months ago I hired a cleaning company. Their estimate was off. I already knew what the job needed. I didn't correct them. I wanted to see what happened when their estimate met reality. Read on LinkedIn →
We finished the job 8 hours late. I went back to the data. Rechecked the formula. Everything looked right. Ran it again on the next job. Wrong again. Read on LinkedIn →
Can you build a company with your spouse without it breaking your marriage? Most people say don't do it. We did it anyway. Here is what makes it work. Read on LinkedIn →
Two things happened recently that every AI PM should sit with. The Claude system prompt leaked. Claude had downtime. Same lesson. Different surface. The model is not the moat. Read on LinkedIn →
Clients describe their home the way it feels to them, not the way it looks to a cleaner walking in cold. That is not a data problem. That is a perception problem. Read on LinkedIn →
How do you know the five-star review you're reading is real? We built a review layer where only verified, completed customers can leave feedback — and made it tamper-proof. Read on LinkedIn →
Most founders research the market. I ran it. 170+ jobs. Both sides of the marketplace. That's not a methodology I'd recommend to everyone — but it's the only reason the product we're building actually fits the problem. Read on LinkedIn →
The problem started at home. Literally. My wife and I are immigrants. Two kids. No extended family nearby. Both of us working full-time. Read on LinkedIn →
25 posts and growing. Follow on LinkedIn for new posts every week.
Follow on LinkedIn →