Siim Haugas

Solo founder from Estonia. Dad of 3 boys. I do not have neat founder mornings or deep-work afternoons. I ship between everything else. Between calls. Between errands. Between family logistics. No office. No team Slack. No standup calls. Just me, the code, and customers who pay. The tools are so good now that one person can do what took a small team 5 years ago. The hard part was never building. It was always selling.

my claude code subscriptions all maxed out, and I have openai subs as a backup. but there's one problem, though. it says that 6.1-sol is in its capacity and says to select another model. ugh

april: gave up on this game. october: 4 weekly opus limits later, it's done.

BRRRT. my A-10 Cuba! itch from 1997, finally scratched. a10canary.com try it out and leave your comment what you think!

same afternoon, two clocks. one call: put a jacket on, i'm almost there, we're at the shop. another: the sales desk has to be up before end of day, someone already promised to show it. family logistics and a deadline, back to back. that was the day.

low-key ported the ios app to android with opus 5.5. it took some effort tho. mostly setting up play store.

thinking about a free, lighter clipper. windows and mac, bring your own keys, public on github. side project of a side project, next to the main one (voxconvo). not a fancy interface. a fast way to cut, because we already know where every word sits. the main product stays the product. this is just the part people would use anyway. but it is just a toy vlog for me where i post stupid stuff that i won't post to X 😅

in my client's project, i ran the evals. honest read: the models got so much better that most of the gain is the model, not us. highest context we saw was 102k tokens. almost everything already fits. the output is precise. what is left is the emotional part. that is the bottleneck, and the evals were useful mostly because they showed it.

+1 joining the trend 🤔 added it to superscribe.io too

half true. tiny personal tools, yes. normies are already building software without even realizing it. but anything a bit more complex takes a ton of tokens, way past what the $20 plan gives you. the people saying it's easy are on $200 plans, often several, and they're in a bubble. and building is only the start. you also have to maintain it, with your own time instead of doing something else. most people would rather pay $20 a month for a polished tool that keeps getting updated. not research, just my take. those big plans are subsidized too. pay API prices and replacing all your tools gets expensive fast. instant software on demand is still a long way off.

oh cool, X has calls now, like google meet. so i added X call detection to superscribe desktop. when a call starts, it offers to record it and transcribes the whole thing. diarized, no bot.

hired 9 people. all of them are me. building superscribe solo.

i was testing the android app that morning. registration had to work the same way as ios. the agent said install chrome so the login would work. i told it that is insane. i do not want to install extra software so our software works. then a survey firm called while i was in a restaurant. i said no, i do not want to talk, and i do not have time. they asked about next week anyway. i said no again, nothing booked, no callback, over in a minute. an hour before that call, wife was waiting at the library main door. she needed the bags, and i could not drive them up. a bus stop sits in the way. so the bags come on foot, or they do not come. son was not ready for football practice. getting him ready was still on me, and then the walk. next week was never going to fit in that day.

you think skipping the gpu means agents are cheap on your laptop. i thought that too. inference from the provider, no local model, should be light. then i ran tens of agents in parallel on my mac mini. the system drained. saas services on the same box started dying. my transcription service got so slow it was lagging. sometimes unresponsive. so i bought a new box and installed linux on it just for the actual services. i don't even run any dev agent on that machine. why does this still happen if the gpu is someone else's problem? because the swarm still lives on your cpu. processes, memory, network, disk. all of it, all at once. if the playground sits on the same box your users hit, you find out the same way i did. curious if anyone else had to split hardware just to keep the product alive.

not devs, but the coding part will and should be replaced by ai. a client's employee sent a vibecoded dashboard. it sits on vercel, with a bunch of packed files. the question was whether it can be integrated. one email, low effort, forwarded like the hard part was over.

what kinda UI is this?? i just want to exit sandbox mode and switch to another account 🙄