DryMerge (YC W24)

Our average user opens our app 0.3 times per day. That's exactly what we want. Here's why: — Most SaaS founders obsess over daily active users. More opens = more engagement = more value, right? Wrong. At least for us. — When we started DryMerge, I tracked every metric. Daily opens, time in app, clicks per session. Then I realized something: The customers using our app every day were the LEAST happy. — Our best customers almost never log in. They set up DryMerge, then forget we exist. Their CRM just... stays updated. → Every email logged automatically → Every call transcript processed → Every follow-up task created → Every field updated correctly Zero clicks required. — The problem with most AI interfaces: They're competing with how you already work. Every new dashboard is another tab to check. Every notification is another interruption. Every approval flow is another bottleneck. We measure success differently now: ❌ Time in app ❌ Daily active users ❌ Click-through rates ✅ CRM records updated ✅ Hours saved per user ✅ Days since last manual fix — If you're still manually updating your CRM after every call, we should talk. Your time is worth more than data entry.

Startups are so damn hard. Last Tuesday at 2:30 AM, a customer went down. Their Salesforce admin had turned on 2FA for our integration user. It completely bricked our connection and prevented us from making any CRM updates for hours. Code red. It took me until 9am to debug, at which point I jumped into nonstop sales calls on 2 hours of sleep. This is what building a company actually looks like. — Not the million dollar round. Not the TechCrunch feature. Not the YC badge. I have all those. They don't matter. What matters is understanding that every real CRM is a Frankenstein monster. - Custom fields duct-taped to legacy systems. - Workflows written by people who left 5 years ago. - Integrations that nobody remembers installing but everyone's afraid to touch. Our AI agents can’t *just* update Salesforce or HubSpot from customer interactions. They need to navigate a maze without breaking the house of cards that somehow keeps businesses running. And that’s *so* hard to build. — When I have tough weeks like this, I think about something our YC group partner Jared said: "I used to be happy whenever I ran into something really hard, because I knew the competition would have to do the same thing." He’s absolutely right. There's joy in struggle. — Every day we eat glass. But it's getting easier. Not because the problems get smaller. In fact, they’re getting bigger and harder. We're just getting really good at eating glass. — At the end of the day, we will deliver AI agents that update your CRM for you. The data will be accurate, and you will love the product. It doesn't matter how much glass we have to eat. If you're interested, let's talk.

My team locked me in a meeting booth for 2 weeks straight. 19 sales calls on Tuesday. 12+ every day after. I was living on takeout and green tea. Here's what 100+ sales calls in 14 days taught me: — 1. After 40 calls, every prospect sounds identical. Everybody has the same pain, the same complaints, and the same story. By minute 6, you want to scream "I KNOW, I'VE HEARD THIS 40 TIMES TODAY." But the prospect hasn’t told the story 40 times. They've been living with this problem for years, finally saw something that looks like it could help, and want desperately to feel understood. The moment you treat call 40 just like call 35 is the moment you stop crushing deals. — 2. Your demo needs to be so good it sells itself. In April, our demo was garbage. Nearly 100% of companies ghosted after the first call. We spent 2 months rebuilding it completely. Now every call ends with some version of "Holy shit" and "This is the future". When they interrupt you mid-demo, say “let’s schedule with my VP for tomorrow”, and shrug off a price you’re afraid to quote, you have a good demo. — 3. If you're building something people want, the pain will be so obvious it's stupid. Every one of the 90 sales teams I talked to had the exact same problems. CRM data is garbage, reps waste 2+ hours on updates, forecasts are complete fiction. One VP told me: "We'll do our sales plays in Google Sheets because Salesforce is so fucked up." This has been life in sales since ‘05. How is it still a problem in 2025? Doesn't matter. It is. The problem is *so* obvious. And so is the solution. — I’m still drowning in calls, but mostly implementation and customer success. This cohort of prospects are closed & happy. I’ve got slots open. If you’re dealing with a nightmare CRM, let’s chat.

GPT-5 is WAY better on our internal evaluations. For real agents doing real work, it's a big leap forward. We threw it at our hardest CRM automation cases at DryMerge. Here are the 5 biggest improvements we saw: 1/ Exhaustive searching When searching, GPT-5 tries every angle. Different spellings, partial matches, indirect connections. It searches like someone who actually cares about finding the answer. Even if a customer uses their personal email or a company has multiple domains, it finds them anyway. – 2/ Handling ambiguity The real world is messy. Edge cases have edge cases. Where's the line between "priority" and "nice to have"? GPT-5 makes sense of it instantly. Custom fields that mean different things in different contexts? It figures out the pattern. – 3/ Following complex instructions We tested it with 200+ business rules, many with exceptions, conditions, and context-dependent logic. Not only did GPT-5 follow them all, it understood the nuanced hierarchy - when to apply exceptions, which rules take precedence in specific contexts. Without being told. – 4/ Learning from context We threw it into situations with custom business logic that would typically require hours of configuration. It figures out the patterns from existing data alone. It can look at hundreds of objects, analyze custom fields, and understand how businesses actually work. – 5/ Extreme reliability o3 was already reliable, but for real production AI agents, 99% is often not enough. Even in the rare, tricky cases that o3 struggled with, GPT-5 is MUCH more consistent. — The craziest part? We didn't have to change our prompts. We just swapped the model and watched our success rate jump. Six months ago, we were burning money on every customer because the AI needed so much hand-holding. Today, it just... works.

Patches O'houlihan in Dodgeball, said it best. "If you can dodge a wrench you can dodge a ball" Hilarious in itself...but it was the "5 D's of Dodgeball" that was the catalyst to this post: 1. Dodge 2. Duck 3. Dip 4. Dive 5. & (you guessed it) Dodge... again... hahaha Me personally? I find that scene hilarious. So it made me think... Can I come up with something equally as dumb (genius) for DryMerge and have the internet voice their opinion? Perhaps?... "Perhaps not" [White Chicks] <--- Another comedy classic This is what I came up with: 1) DryMerge Destroys Dirty Data. Dedupes. & Directs Data to Drive Revenue 2) DryMerge Does Data Diligence. Driving Dollar Defining Deals 3) Dirty Data Dragging Do Down? DryMerge Drops Doubt. Drives Delight 4) Ditch Dirty Data. DryMerge Drives Dependable Deals. and last one 5) ...DryMerge Destroys Dirty Data. Dedupes. & Directs Data to Drive Revenue Number 3 was so FORCED. But I was having fun with this exercise. I laughed and said wth. So I shipped it. Which one was your favorite? --------------------

Anthropic just dropped Claude Opus 4.1. OpenAI released their open source models 34 minutes later. It's gonna be a busy week... For now, Opus 4.1 is the best coding model in the world. Windsurf says "It's a full standard deviation better than Opus 4". 74.5% on SWE-bench Verified, slightly edging out their previous score of 72.5%. Just 7 months ago, the best models were <40%. — I've been testing Opus all morning. It has already successfully fixed 2 bugs so far that previous models couldn't solve. It definitely seems more optimized for Claude Code than Cursor, but this could change quickly. It's much better at searching large codebases. And its edits seem more precise: it touched exactly what needed fixing and nothing else. GitHub, Rakuten, and Windsurf are all reporting massive improvements. — Meanwhile, OpenAI just open-sourced gpt-oss-120b and gpt-oss-20b. Their first open language models since GPT-2. Apache 2.0 license. Full chain-of-thought exposed. Tool use built in. The 120B model hits 96.6% on AIME 2024. That's competitive math performance in an open model you can run yourself. — GPT-5 is rumored to come later this week. And Google DeepMind just hinted that they're releasing a new model too. Here's what's actually happening: What cost $1000 last year costs $10 today. What was API-only yesterday is open source today. What took GPT-4 weeks to code takes Opus minutes. — We're watching the most competitive technology race in history unfold in real-time. Every lab knows that being #1 for even 48 hours matters. Every pricing cut forces everyone else to match. Every capability leap becomes table stakes within months. You get to ride every improvement. Switch models with a config change. Build features that were impossible last quarter. While the labs fight for supremacy, builders get superpowers. What are you shipping this week?

61 Meetings booked in 24 hours by my CTO. As THE Lead Generator at DryMerge. I'm so cooked. Two weeks ago, our CTO, Sam Brashears went on a 3 day LinkedIn post bender. One viral post after another. It was like Steph Curry sinking 3-Pointers from the logo. He landed us 61 meetings in 24 hours. You read that right, 61!! 70% were highly qualified leads with the same hair on fire problem. This didn’t just happen by accident though… It was carefully planned. One week prior to this unfolding, we had an internal strategy session to diversify our outreach outside of traditional channels (Email, Calls, InMail). Posting good content on LinkedIn as frequently as we could, was our number one priority. - But why? Leveraging social media and tapping into a space that would place DryMerge’s value prop to accurately update customer CRM’s across their existing revenue stack, was top of mind. And…It worked. Maybe it was luck. Maybe it was skill. (Sam humbly believes it was both) Call it what you want. What we do know is that we now have the data and pipeline to show for it. - Here is what we took away from this beautiful chaos that ensued: 1. Work smarter, not harder 2. Do not be afraid to post. Might just be the post that adds + $1Mil in pipeline overnight 3. The posts that yield results haven’t always been about what DryMerge does. Diversify. 4. Enjoy the process. Enjoy the journey. Keep going. Oh yeah. And the whole “I’m cooked” thing? Meant to say “I’m cooking”. 62 more meetings overnight coming right up 👨🍳

OpenAI has 5 days to launch GPT-5. Not because they want to. Because the EU is forcing their hand. August 2nd: EU AI Act requires all new models to expose training data, architecture details, and safety measures. Enforcement? €35 million or 7% of revenue for non-compliance. That's *billions* for OpenAI, which is already running at a loss. OpenAI's racing to launch before the regulatory walls go up. — So GPT-5 likely drops *this week*. Here's why it matters: Reasoning models like o1 and o3 solved instruction adherence. At DryMerge, our agents went from 70% to 100% success overnight. But agents needed *perfect* prompts to hit those metrics. GPT-5 solves *intent*. Sam Altman calls it "magic unified intelligence" - one model that automatically adapts to what you need instead of making you choose between models. No more switching between o3 for reasoning and 4o for speed. The model figures out which mode you need with *excellent* taste for writing, code, and agentic tasks. — This is huge for anyone building in AI. Why? Because customers don't give perfect instructions. They say "sync my emails to Salesforce" when they mean "extract the contract value from this forwarded thread buried under five signatures." o1 needed us to translate. GPT-5 gets it. It's the difference between an intern who needs a detailed spec and a senior contributor who already has years of intuition. — Every model release so far has made our product materially better, but there's some buzz that this one might not deliver the same results. "Just a router". I trust in OpenAI, do you?

2 weeks ago, I started posting on LinkedIn. Since then, I've gotten 2.3 million views and over a million in pipeline. Here's exactly what happened: 2 weeks ago, my co-founder and I had a problem. We wanted more leads. Fast. I'd never posted on LinkedIn before. But I made a commitment: post every weekday for 90 days. Day 1: I sat down at my desk and spent 5 hours on a single post. 5 hours! I rewrote the hook 20 times. Edited the image padding for 10 minutes. Obsessed over every word. Then I hit publish. — It went viral and got 160k views. So I did it again the next day. And the next. Each post took hours. I treated them like mini-essays. Crafted stories about our journey that I'd never shared publicly. The results? → 2.4M views → 9,000+ likes → 3,000+ new followers → 71 demos booked → Over $1M in pipeline — But here's what's crazy: only ONE of those posts actually mentioned what we do. Last Friday, I wrote "OpenAI just killed my startup." I told the story of how we pivoted from general automation to the one thing sales teams actually hate: updating CRMs after every email, call, and Slack message. How sales reps waste 2+ hours a day on data entry. How every CRM is a "beautiful disaster" of custom fields and broken integrations. That single post generated 90% of the demos (61)! My co-founder's calendar? Completely booked. Wall-to-wall demos since Friday. Turns out people really hate updating CRMs. — Now it takes me <90 minutes per post. Still slow. Still obsessing. But getting faster. I wish I had some clever growth hack to share. But honestly? I just spent way too much time making each post as good as I could. Turns out that works.