What's possible

Concept work runs alongside the client work here: things built to find out how far something goes. Two of them below.

trydotto.live

The precursor to an app in the store, built in the browser because everything that gets expensive after submission can be answered here: how the environment feels, which interactions people actually accept, where they drop out. What holds up here goes into the app; what doesn't cost nothing but the week it took to build.

The shape it runs in: a dot on a screen. Every so often it goes hollow, leaving a couple of seconds to tap. A tap at any other moment, one missed dot, or leaving the screen ends the run. Nothing in it distracts — the phone it runs on already does that.

• The engine:Pure TypeScript with the clock and RNG injected, so a whole run is deterministic and testable without a browser.
• A test inside the test:Half of all sessions never get asked to predict how long they'll last. The arm is sticky and rides on every event, so the funnel splits by it and the mechanic has to earn the friction it adds.
• Analytics with nothing to consent to:No cookies, no device storage, no person profiles, EU host. The Meta pixel is geo-gated at the edge and never loads in the EU, so the page needs no banner.

Automated content pipeline

Running for a niche account on Instagram. It mines what the niche actually struggles with, writes in one fixed voice, renders the post or reel, blind-judges its own output, and only then publishes. The same pipeline runs any number of accounts — each one a folder with its own voice, niche and credentials.

The framework underneath is deliberately plain, and that is the point: how realistic and how fine-grained the output gets depends on the effort put in, not on the architecture. This one was not pushed any further than it had to be — the ceiling is a long way up.

• Mine real demand:A miner pulls threads from the niche's subreddits and an LLM distills them into a pool of concrete pain/wish signals — topics come from what people actually say.
• Voice is the moat:The writer is locked to a per-account system prompt. The output isn't "an AI post" — it's one consistent character, and that voice is the whole defensibility.
• Render headless:Carousels are composed in headless Chromium from templated layouts; reels are assembled from a generated script — TTS voiceover, cued stock b-roll, captions, cover — into a 9:16 MP4.
• Judge before publish:A blind multimodal judge scores the finished media against a rubric and is the gate: below threshold it is discarded, never posted. Quality control with no human in the loop.
• Autopilot:A daily planner scatters publish slots; a per-minute ticker generates → judges → publishes just-in-time, so posts land on a human-looking schedule, not on the hour.
• Learn from results:A collector pulls each post's Instagram insights — reach, saves, shares — and scores its real pull. A weekly pass re-weights which formats get written, biases publish times toward the hours that land, and hands the writer a brief of what's working plus fresh angles. The account tunes itself on its own numbers.
• Multi-account by design:Account is a per-request parameter. Adding the next one is a folder — its own prompts, creds, and data — with zero pipeline changes.

The LLM call is the easy 10%. The job is the voice, the render pipeline, and a judge strict enough to let it post unsupervised.

Both of these are examples, not a ceiling. If something surfaced while reading — an idea, a process that eats a day every week, a thing nobody is sure is even possible: let's talk about it.