Levon Terteryan

I think a lot of AI assistant UX is going to come down to boring transport details, not model IQ. The recent OpenClaw Telegram threads made that pretty obvious to me. On paper, partial streaming sounds good enough. In practice, people are running into duplicate replies, visible NO_REPLY leaks, and all the weirdness that comes from faking real-time chat by constantly editing messages. At the same time, the community is already asking for native Telegram streaming now that the API supports it more cleanly. That matters more than it sounds. If your assistant lives in chat, every flicker and duplicate makes it feel less trustworthy. The interesting part is that this is not a frontier-model problem. It is product plumbing. Better transport makes the same agent feel calmer, faster, and more competent on your phone. I keep coming back to that with OpenClaw: the useful breakthroughs are often in the orchestration layer, where a small fix changes how human the whole system feels.

I don't eat meat or fish (18 years) I don't eat dairy (7-8 years) I don't smoke (never smoked) I don't drink (quit 3 years ago) I don't sleep in (very rarely) I don't take shortcuts I don't miss daily workouts I don't lament or complain I don't stop

We had a waitlist on Playgram for about a month. Felt like the responsible thing to do, you know, control the flow, make sure everything works before letting people in. The usual founder reasoning. Then we turned on paid ads and I realized how dumb that was. People are clicking on our ads, landing on the site, and the first thing we ask them to do is... wait? That's not onboarding, that's a rejection letter with extra steps. So on a Friday call I just told the team, let's remove it. Even if not everything is perfect yet. Sometimes the best product decision is just getting out of your own way :)

I think the most important OpenClaw browser upgrade is not that it clicks buttons faster. It is that the browser can finally stop feeling fake. A lot of browser automation breaks down the moment you leave the sandbox and hit real life: your Gmail session, your logged-in Notion tab, the weird internal tool that only works in the browser you already use every day. The new live Chrome attach flow matters because it shifts the agent from a demo environment into your actual working context. One toggle, real tabs, real cookies, real logins. What makes this interesting is that the community pain around OpenClaw browser control was never really about wanting more automation. It was about reliability and setup friction. If the agent can work inside the browser you already trust, a bunch of awkward relay rituals disappear. That feels like one of those small infrastructure changes that quietly makes the whole product much more usable.

I've been trying to describe what Playgram does for months now and I keep rewriting it. Every pitch deck, every landing page, every time someone asks "so what do you do?" at a conference, I say something slightly different. Yesterday I typed out yet another version in our marketing channel: "Playgram is a team workspace that unifies all AI models in one interface with an infinite shared memory and context." And for the first time it actually felt like the right words in the right order. Not too technical, not too vague, not trying too hard to sound impressive. The funny thing is it's basically what we've been building the whole time, I just couldn't say it properly until now. Probably says something about the product too, that it took this long to find one sentence that explains it :)

So I had this CRM sitting on Amazon that I've been meaning to move for months. It worked fine. "Fine" being the word you use when something technically runs but you've stopped telling anyone where it's hosted because you know they'll ask why. Yesterday I finally sat down to move it to Railway, and the database from Postgres to Supabase Postgres. The thing is, I didn't actually do it myself. I gave it to Codex and it moved everything, not just the data but all the user logins, passwords, the whole thing. Took maybe 20 minutes. And now I'm sitting here thinking, if an AI can migrate an entire CRM with auth credentials intact in the time it takes me to make coffee, maybe we should've done this six months ago. We had a lot of very serious infrastructure discussions that could have been one afternoon :)

We had a real debate inside the team yesterday about authentication. Sounds boring, but it's one of those decisions that touches everything and you can't easily undo later. The question was whether to use a managed service like Supabase Auth, or go with Better Auth and self-host the whole thing. My gut reaction was to self-host. My reasoning was pretty simple, maybe too simple: since AI is going to be writing most of our code anyway, the maintenance burden isn't what it used to be. Just let the agents handle it. Our lead engineer had a different take though. He pointed out that even if AI writes the code, someone human still needs to own the security policies, review the critical parts, and keep up with OAuth provider changes. "Building it is one thing, but we're the ones stuck maintaining it," he said. He's right, obviously. The part I keep underestimating is that AI doesn't remove the need for human judgment on security, it just changes what kind of judgment you need. We decided to go with Supabase as a database and auth provider.

One of the things I spend a surprising amount of time on as a founder is pushing people on the team to actually use AI. Not just have opinions about it. To use it. The ironic part is that we're an AI company. AI-first in every external conversation we have. And still I find myself in this same conversation with the team pretty much every week. Someone explaining why this particular tool isn't quite right for them. Why now isn't the best time to try something new. Why their situation is a bit different. All reasonable. All a little resistant. I don't think it's about the tools. It's just human nature. Every new thing meets the same friction, and part of my job is to keep nudging people past it. Every role, every function, the whole team. Nobody really tells you this part before you start a company :)

Hired a new developer this week and I realized how weird onboarding has become. We used to do these elaborate first days with structured training and shadowing sessions. Now I just give them access to the repo, the Slack, the Cursor subscription, and say "just start working." Roberto has been with us for three days and he is already pushing code. I keep waiting for someone to tell me this is irresponsible but honestly it is working better than the two-week orientation we used to do. He figured out our JSON structure by himself and built a deterministic script that splits files into readable chunks. Maybe the best onboarding is just trusting people to figure things out. Or maybe we are just too small to afford proper process yet. Probably both :)

We got a response from a GCP partner about cloud credits yesterday. Without VC funding they'll give you up to 25K, with funding up to 200K. They start you at 5K and work up from there. No commission, which is nice. But then we started actually comparing platforms, and it got complicated fast. AWS Amplify vs Google Cloud Run vs Railway. Our engineer had already done a comparison doc and the verdict on AWS was basically "high complexity, requires DevOps expertise, bill shock potential, over-engineered for this stage." Which is a polite way of saying it's a lot. So the plan is to start on Google Cloud Run, apply for credits in parallel on both GCP and AWS, and if the credits don't come through or something doesn't work, just move the Docker container somewhere else. That's the beauty of containers I guess, your hosting decision doesn't have to be a marriage. More of a month-to-month lease with an easy move-out clause :)

We've been testing Playgram on slow 3G this week. Not because our users are on 3G, but because if it works there, it flies everywhere else. Found some embarrassing bottlenecks. The kind where you wonder how nobody complained yet. Fixed the worst ones. There's a trade-off though. Image generation got slightly slower, from about 3 seconds to 4-5. We could build a whole speed-detection flow to handle this dynamically, but that felt like exactly the kind of over-engineering that creates more problems than it solves. So we're shipping it as-is. Fast for most things, slightly slower for images. Honest about the trade-off. Better than pretending everything is fine and hoping nobody notices :)

@oliverbrocato this is so cool Oliver!

We raised $64M for this moment: Introducing Bustem. Bustem scans the internet to find and eliminate 100% of counterfeits RT + comment “SCAN” and I’ll send you a list of every scammer targeting your brand 🫵

We interviewed a ton of senior developers last week. Did a lot of calls back to back, which was honestly a bit exhausting, but also kind of eye-opening. One of them had been freelancing for the last six months. Before we even got on the call, he had spent three hours going through our app on his own and shared a few security optimization ideas! He had not been hired yet. He just did it. I have a much clearer sense of what I am looking for. Someone who has agency and just gives a shit before they even get paid to. We will see how the rest of the process goes :)

Ever seen a gym at an airport? :)

My CMO who is the only person running our marketing efforts told me he don't need any humans for the marketing team :) We set up OpenClaw, which is an open-source agent framework, on a $32 a month server. It connects to Ai providers as well as different tools like Slack, Telegram, GitHub, Gmail, Notion, Google Analytics, pretty much everything we use day to day. The agents have memory, they can browse the web, execute code, transcribe meetings, send messages. They work while we sleep. We spun up 6 agents with openclaw: 1. One main one who orchestrates everything 2. One front-end engineer (all the work on marketing website, code changes, pushing to github and deploying to vercel) 3. One back-end engineer - all kind of automation stuff 4. One content-manager - helps with social media post ideas, scheduling etc. 5. One outreach/ SDR agent - cold outreach over email and linkedin. 6. One HR assistant - connects to the recruitment website where we post our vacancies - can post new ads, check applications, download CVs to google drive, add them to a google sheet, score them for sutiability, create a task in Asana. It's not super easy and straighfotrward to set up Open Claw - It is not plug and play. But once it was running, things that used to require back and forth with multiple people just kind of happen on their own now.

We were using OpenAI Whisper for voice transcription in Playgram but it wasn't great - languages (if not English) would get mixed up, transcription wasn't always accurate, so we switched to Deepgram a few weeks ago - faster & noticeably more accurate.

I never used the vocabulary of someone working for me. It's always someone working with me, and maybe that's one of the reasons that I have people working with me for many, many years. My CFO, for example, has been working with me for 23 years now. She started as an accountant in one of my previous companies in a different country where I lived before, and then she became chief accountant. Then she was also managing my real estate and also being the CFO, and now she is the CFO for all the current companies that I run.

We had this annoying bug this week where DOCX, PPTX and a few other file formats were taking around 10 seconds to process instead of the usual 2-3. Not a huge deal on the surface, but when you are using an AI product and it just sits there for 10 seconds after you upload something, it feels broken. Turned out it had nothing to do with the AI models themselves. It was a format compatibility issue with certain LLM providers. Some files were not being parsed cleanly and the whole thing was just grinding. The fix was pretty simple actually. We now convert everything to PDF first before sending it to the model. Full compatibility, tables come through properly, and it does not break when we switch providers. Processing time went from 10 seconds down to 2-3.

Had a call with Gaingels yesterday. They are one of the larger venture groups out there, something like 5,000 investors across 2,800 portfolio companies. It was a pretty relaxed conversation honestly. Their process is 3-4 weeks, minimum check around $100k, average somewhere between $250-500k. They like companies that are somewhere between seed and Series A with a bit of validation already in place. What I did not expect was how much they cared about things beyond the money. AWS and Google Cloud credits, talent pipelines, intros to other portfolio companies. More of a network play than just capital. We are not raising right now. But I have learned that the conversations you have before you need money are usually the most useful ones :)