I shipped my first AI-built game in 7 days. Stardust Drifter: every star can kill you, or slingshot you to the next one. First playable: 15 minutes. The rest of the week: real playtests, AI agent tests, same-day fixes. Play (best on desktop): dabrodev.itch.io/startdust-d…

$0 infra at ~100 simulated users. Curious what your web app stack looks like? Vercel is great, but once you scale Server Actions, CPU limits hit fast and force upgrades. Supabase is solid too, but Storage and Edge Functions can get expensive quickly. Here’s what I switched to: → All server-side logic → Cloudflare Workers → Storage → Cloudflare R2 → Database → still Supabase (for now) I’m also considering moving the DB to Neon, especially now that it supports auth. Current setup: ~100 agents simulating users + organic traffic Last 30 days: → 2.4k unique visitors → 97.5k requests → peaks ~10k/day → ~61% cached So far: $0 in infra costs.

$0.20. That’s all it takes to generate a 10s AI talking head. Image → $0.03 Voice → $0.02 Video (with lipsync) → $0.15 Face + voice fully generated. 480p is enough to test fast. 720p? 4× more expensive per second or just upscale. Want the workflow? Comment “video” 👇

Most remarketing is just repetition. Same product. Same message. Same pressure. But retargeting should feel like: “Hey, we met before.” That is why Livra builds sequential remarketing as a story arc. Act I: remind Act II: prove Act III: close

Nano Banana 2 vs Nano Banana Pro vs GPT Image 2 What do you think?

Most UGC ad generators ask you to choose the actor. But how should you know which actor will actually sell? You don’t. You test. Same product. Same offer. Same angle. Same format. 3 different actors. 1 ad set. Only one variable changes: the face. Then Meta picks the winner based on performance, not opinion. That is the difference between generating ads and operating a campaign loop. Modern ad software should not just give users more choices. It should guide the test, find the winner and turn that winner into the next step.

You don't need Hermes Agent to run ads fully autonomously. You need an autonomous campaign loop. Because the problem with ads is not that founders cannot generate one more creative. The problem is that they do not have a system for what happens after the creative is generated. Most campaigns fail here: too few angles tested, too much spend burned on losers, no clear winner, manual decisions made too late, remarketing that repeats instead of extending the story. That is why an “AI ad generator” is not enough. A performance campaign needs an operator. A system that can: read the landing page, build the brief, generate variants, launch the test, monitor performance, pause weak creatives, scale the winner, and trigger remarketing automatically. That is what I am building with Livra. Current campaign structure: 4 creatives 2 formats 1 ad set 24/7 monitoring automatic loser pausing winner scaling sequential remarketing The point is not to create more ads. The point is to create more signal for the algorithm and remove manual campaign babysitting from the founder’s day. For the user, this changes 3 things: 1. Faster testing You are not waiting days for new creative iterations. 2. Less wasted spend Weak variants are paused instead of soaking up budget. 3. Better follow-up The winning creative becomes the base for sequential remarketing, not just a one-off ad. Most AI ad tools stop at creative generation. Livra is built to run the campaign. I created a full guide showing the exact autonomous ads loop: test → learn → pause losers → scale winner → remarket. Like this post, comment ADS, and follow me so I can DM you the full guide.

Livra + Claude Opus + GPT Image 2 = Autonomous Ads Manager Phase 1 in Livra is a pure casting test. Three different actors, three real Meta-feed creatives, one shared product story. The only variable is the face. Meta optimizes inside the ad set and the winning actor automatically carries into Phase 2 scaling and Phase 3 sequential remarketing.

Another version with the new gpt-image-2

Today I rewrote how Phase 1 cold testing works in Livra. Old version: one persona rendered through three different style approaches. Same face, three vibes. New version: three different actors from my persona library, one per approach. Different faces, same product. The reason matters more than the change. Cold testing in Meta is signal extraction. You want one clean variable so the algorithm tells you exactly what worked. Most ad generators rotate styles or copy or hooks, often two or three variables at once, and the learning phase becomes guesswork. If the variable under test is the actor and only the actor, the answer Meta returns is unambiguous. This face sells your product. That answer compounds downstream. The winning persona scales in Phase 2 with the same face plus minimal variations. Phase 3 sequential remarketing rests on the audience recognizing that face from cold test. Casting decision in Phase 1 becomes the spine of the whole funnel. Quality of test setup determines quality of every later phase. That is the lesson. Image rendered through Vertex AI, 4:5 portrait, 2K. Same prompt on both: NB 2 (gemini-3.1-flash-image-preview) takes 32 seconds, NB Pro (gemini-3-pro-image-preview) takes 44 seconds. #buildinpublic #ai #saas #metaads

It depends on the app. In one of my AI Influencer apps, the AI persona has a lifecycle, so it wakes up, goes to sleep, chooses different activities, checks trends, and makes posts. It’s fully autonomous, but some activities can also be triggered manually. So to give the user more control and overview, I added a lifecycle log and a real-time status showing what happened and what is happening right now.