Tomas David Ye

We stopped celebrating recommendation accuracy. We started measuring overrides. That changed almost every conversation we had with warehouse operators. A model can recommend the right carton 99% of the time and still fail if people don't trust it enough to use it. An override isn't always resistance. Sometimes it's a broken process. Sometimes it's outdated warehouse logic. Sometimes the operator knows something the system hasn't learned yet. That's why I don't think override rate is a failure metric. It's a learning metric. Every override is evidence that the software and the operation disagree about reality. The job isn't to eliminate every override. The job is to understand which ones should disappear and which ones should teach the system something new. The fastest way to lose adoption is to treat every manual decision as user error. What has your team learned from the decisions people refused to let the software make?

We built an AI onboarding assistant. Answer any question, day one, no waiting on a trainer. Turnover in the first two weeks didn't move at all. We sat in on floor inductions to find out why. New hires weren't asking the assistant questions. They were asking Marisol, a picker three years in, because she'd answer with "yeah, and also" and tell them the thing that wasn't in the manual. Which bathroom actually has paper towels. Which supervisor to avoid asking twice. That the AC in bay 4 only works if you don't mention it's broken. The assistant answered questions correctly. Marisol answered questions and told them they were going to be okay here. We stopped trying to replace her and started paying her for two extra hours a week to do onboarding walk-throughs, with the AI handling only the paperwork behind her. Two-week turnover dropped 30% in the next quarter. We'd built a very good answer machine for a trust problem.

People ask why we spend so much time inside customers' warehouses instead of just shipping the software and moving on. honestly it's because that's where we find out what we got wrong. you write the roadmap in a conference room and it sounds airtight. then you're standing there during a shift change, watching someone try to use what you built, and you realize you never accounted for the fact that shift change is chaos, everyone's rushing, nobody's reading the screen carefully. You don't get that from a feature request email. you get it from standing next to the person actually using the thing at 6am when they're tired and just want to get through the shift. We started calling ourselves "forward deployed" somewhat jokingly at first, but it stuck because it's true we don't think go-live is the finish line, it's honestly where the real learning starts. I'm not trying to prove the software was right when we built it. most of the time it wasn't, not entirely. i just want to leave every deployment knowing more about how the warehouse actually runs than i did walking in. curious what's something you only figured out after actually being where the work happens, not before

For the past few years, we've been building quietly. Visiting warehouses. Talking with operators. Sitting down with founders. Learning from every deployment, every customer, and every mistake. Most of those conversations never made it online. We're changing that. From today, you'll see a lot more of what happens behind the scenes: warehouse tours, customer visits, industry events, product decisions, and the realities of building AI for logistics. We're starting at Fetch Fulfillment, one of our earliest customers. In this first video, we go inside a real 3PL to show how a warehouse actually operates, how orders move from click to shipment, and why the small operational decisions most people never see are often the ones that matter most. This is the first of many. Watch the full video on YouTube. Link is in the comments. We'd love to hear what you think.

We crashed our own demo in front of a prospect because one SKU in their test data had negative dimensions. Not missing data negative. A box logged at roughly -4 x 12 x 8 inches, sitting in their warehouse management system for years, quietly excluded from every report that should have caught it. Our system choked on it live, on a call, in front of the person deciding whether to sign. We didn't talk our way out of it. We showed them the SKU, showed them how it got there, and told them their existing tools would never have caught it because nothing in their stack was built to expect an impossible number. They signed anyway for exactly that reason. The bug we were most embarrassed by became the strongest part of the pitch.

What if logistics content felt like a documentary instead of a presentation? That’s the direction we’re heading. Every frame should teach something. Every conversation should challenge something. Every video should leave you understanding the industry a little differently. No filler. No corporate scripts. No recycled AI headlines. Just stories worth watching. Welcome to what’s next.

Week one: 40 signups, I was on top of the world. Week four: 40 signups, still active, and I finally understood the difference. Signups feel like progress. They're a raffle ticket. Retention is the actual scoreboard and it's a much less exciting number to post about, which is probably why nobody does. If you're building in public, the boring metric is usually the honest one. Impressions and likes are a nice mood boost. They've never once paid a bill. What's the vanity metric you had to stop caring about?

A few of you asked what's actually changing Monday. Honest answer: I've been asking myself that too, every day this week. Three years of doing this alone meant I got used to a certain rhythm. Write it, post it, see what lands, move on. Simple. This week hasn't felt simple. More people in the room. More opinions on what's worth showing you. More "wait, is this actually good enough" moments than I've had in a long time. Growth does that, I guess. It doesn't just add numbers. It adds pressure to not waste what you built. I don't know yet if Monday will feel like a big moment or a quiet one. Probably somewhere in between. But it felt right to tell you it's been a real week, not just a countdown. See you Monday.

Nine minutes. That's all the time he gave me, standing in his own warehouse, contract already flagged for churn. I didn't pitch. I asked him to walk me to the pallet. Four minutes in, we found it: a SKU his own team had mislabeled three weeks earlier, blamed on "the software" ever since. Five minutes after that, he just stood there, not saying anything, staring at a label his own person printed. He renewed. Six figures. Saved by a plane ticket most software companies would call a bad use of resources. Here's what doesn't fit in a pitch deck: software can send an email when something breaks. It cannot stand next to someone and watch them realize they were wrong about who broke it. We build software. The renewals get saved on the floor, in person, every time someone shows up. When's the last time your vendor actually did?

Our customer's AI acceptance rate dropped from 91% to 58% overnight. Nothing changed in the system. No update. No new integration. No model retrain. Same warehouse. Same AI. Same SKUs. Same carriers. Just a different number on a Monday morning. We spent two days going through logs. Clean data. Clean recommendations. Clean everything. No technical explanation. On day three someone asked a question nobody had thought to ask. "Who came back from vacation this week?" One supervisor. Seven years on the floor. He didn't send a memo. He didn't hold a meeting. He just came back and started doing what he always did. Walking the floor. Watching recommendations come in. Quietly saying "nah" to the ones that didn't feel right. And the team his team followed his lead the way they always had. Not because the AI got worse. Because their most trusted person in the building had an opinion. And his opinion carried more weight than 91% accuracy scores and confidence intervals nobody on the floor had ever been taught to read. We didn't fix the acceptance rate by retraining the model. We sat with him for two days and asked him every question we should have asked before we deployed anything. Which recommendations do you trust? Which ones feel wrong even when the number says they're right? What does the system not know about this floor that you know by instinct? Three weeks later acceptance rate was back to 89%. Not because he changed. Because we finally asked. The most important variable in your AI deployment isn't your model. It's the person your team trusts most. And if that person has doubts you haven't addressed  Your acceptance rate will tell you. Every single Monday morning.