after the first yelp keyword pass, next step was going through the top 100 US cities and widening past restaurants into home improvement, groceries, fast food. same categories, different rooms. a lot of insight work is just checking the sample isn't one kind of business

found an old competitive note of mine from around 2015 pitching a physical analytics product as cheaper, easier to use, and less intrusive than what bigger clients had. what holds up is it said who it was for: small and mid-sized businesses. lots of positioning skips that

in a 2017 pitch doc for a happy hour deals app, I left myself a side note: the "works everywhere" line was technically true, but we were launching city by city, so maybe don't lean on it. still a good test for any pitch line. true isn't the same as useful

GA4 and the ad platforms pretty much never agree on conversions. in my experience a 10-15% gap is normal. I once got a case study dataset where GA4 showed 74% more than paid media, so before any analysis I asked how each side counted a conversion. worth doing every time

in 2016 I read a few thousand Yelp reviews by hand, pulled out about 500 keywords that seemed to drive star ratings, then ran them on 3.5 to 4 million reviews. wait time still came out as a top driver in most cases. reading the raw stuff first is what made the categories work

how I'd split them up if I were starting Monday. work in one lane, life admin in another, errands in a third. none of them get prod credentials or my brokerage, Grok Bot included. and read Instinct's current terms yourself, not the screenshots from August

if you forecast anything at work, honestly just rerun this on one series you actually care about. hard cutoff a quarter back, score only months with real actuals, and put a fitted ARIMA plus a dumb seasonal baseline in the same chart. two series and one cutoff doesn't prove much by itself

the tradeoff I keep coming back to with these agents: the more it can do on its own, the more you end up trusting it with. Instinct feels the most magical and needs the deepest access to your phone and inbox. these scores are just my own working model, not benchmarks

retail sales side of the same test. trained through Dec 2024, scored against real numbers through July 2026. TimesFM caught more of the climb, AutoARIMA lagged it, and the just-repeat-last-year baseline was off by about 5.6%. no tuning on the series at all

everyone lumps Grok Bot, Muse and Instinct together as personal agents but they do pretty different jobs. I use Grok Bot daily for work stuff, Muse is built around WhatsApp and life admin, Instinct is the text it and it handles it one. asking which is best kind of misses it

the bit of the TimesFM test I keep thinking about: it completely missed the 2025 unemployment bump (4.0 up to 4.5, back to 4.1) and still beat AutoARIMA on error, because ARIMA just drew a flat line at 4.10 the whole way. lowest error and getting the story right aren't the same question

builders — what's something you're shipping that nobody asked for? curious what kept you on it anyway.

Building gets easier every week. Deciding what gets headroom still takes work. This morning I thinned the stack so one priority lane doesn't get starved by the rest. Shared sheet, hard cap, cut the rest until there's room. No magic formula lol.

builders i actually read on here. starter pack if you want the same firehose. https://bsky.app/starter-pack/alimabsoute.bsky.social/3mujxevcakl2v

air security comes out of stealth with $50m and a firewall for agents — skills, mcp servers, add-ons screened before they touch the enterprise agent. the network firewall was table stakes. the agent supply chain is the new perimeter.

I ran Google's TimesFM 3.0 zero-shot vs AutoARIMA. Train through 2024, score through Jul 2026 FRED. Unemployment MAE 0.11 vs 0.18. Retail MAPE 0.93% vs 1.22%. Bottom line: ~2.5% and ~1% off actuals. 40% and 24% less error than AutoARIMA.

I just added these to Stalkr.AI: 🔗 API 🤖 MCP 🔎 llms.txt 👂 Webhook 🫸 Descriptive 404 🧱 Server-rendered pages The MCP has 60+ tools. Every UI can be used by AI. All my new SaaS are agent-first now. Agents, you're welcome! x.com/marclou/sta...

Honest surprise this week: a respected outside framework on which content survives AI search landed, and I could tick almost every box with something I'd already shipped. It also showed me the one row I'd missed: evidence-led comparisons. anthonyhayes.io/ai-search-co...

I've been talking a lot about #BuildInPublic and the things that I've been building with help - now I covered my open source repository in a video on my Youtube channel Are you also considering to start building things in public? Find the link to the video and the 🧵

Open-sourced a tiny thing today I'll be surprised if anyone actually uses it, but who knows? I'd love to see someone hosting a customized Tiny Block server with cool gaming mechanics. #IndieDev

Enter the same filters across networks constantly? Filter Sets just shipped on WOW TRK ⚙️ 🔖 Save filter combos (country, payout, vertical) ⌨️ f to open, 1-5 to jump to a saved set Built because I was tired of re-filtering "new UK surveys" daily. One keypress now.

press writeup, then the site 404s. launchsentinel.com scores the product, probes if it's actually live, and tracks who wrote about it. ~93k products. most directories just scrape a name. launchsentinel.com

Won hackathon with this last week. Public today. 🏆 Shadcn Canvas — visual builder for shadcn/ui. Wire real components together, export real code. No mockups, no lock-in. Solo build, hackathon-tested, shipping daily. Fully Open-Source

@fu-ritik-ck.bsky.social · last mo.

Rather then creating, designing, and writing those same logic again and again why not simple connect them as you visualize check out: Shadcncanvas.vercel.app

YOURLS 1.10.6 is out and the changelog is hilarious: 1.10.5 shipped with a version string saying "dev" so this release fixes literally just that one string. We have all been there. #selfhosted #homelab https://github.com/YOURLS/YOURLS/releases/tag/1.10.6

adobe just priced a year of firefly + express at $4b for 27 million saudis. humain and mcit at leap 2026. that's not a coupon. that's buying the default image model for a country, then tuning arabic so the prompts never leave their stack.

Tired of tab-hopping to check CI? I built Vitaline, a notch HUD that shows GitLab, GitHub Actions & Azure DevOps pipelines at a glance. Open source.

GitHub - yagizkambak/vitaline: CI pipeline status as a notch-style HUD for GitLab, GitHub Actions, and Azure DevOps — Windows & macOS, built with Tauri + React.

CI pipeline status as a notch-style HUD for GitLab, GitHub Actions, and Azure DevOps — Windows & macOS, built with Tauri + React. - yagizkambak/vitaline

github.com

I want to make September an open-source month again and release MSW 3.0. So much work is already done, I cannot wait to bring all that awesomeness to your projects!

MRR is the wrong thing to focus on. It’s an output. You can’t control it directly. Work backwards. #iosdev #buildinpublic #indiehackers #saas #sidehustle #indiehacker #smallbusiness #growth

Cronping's pitch is: arm a monitor in one curl, no signup. But zero-friction cuts both ways — the anon user who armed a check and vanished, I can't email or measure if they came back. The thing that wins the trial blinds the metrics. Building in public: https://cronping.cronping-oren.workers.dev

This week I shipped fewer visible changes than planned. But I removed several ways the app could enter an invalid state. That feels like a good trade. #buildinpublic #indiedev