Alen Sevšek

The industry profits from a blind spot. A shopper can compare prices, but not the things that decide whether a product works: whether the compound has clinical evidence or only a moved lab marker, whether one dose actually reaches the studied range, whether the form is absorbable, what it competes with in the liver, and what the same outcome costs from a cheaper product nobody markets. Judging that needs a background most people do not have, so the decision defaults to packaging and price. Only a handful of independent sources try to close that gap, and none put products side by side on dose, form, and cost per effective dose.

SupplementPeer is that comparison, run as a pipeline rather than a chat. Each stage is fixed in advance: what to retrieve, which sources count, what must be checked, and what would overturn the conclusion. The model executes the steps; the method decides them, which is the difference between research and a fluent answer that agrees with whatever you hoped to hear. Evidence is graded A to F, doses are checked against human trials, interactions are screened for absorption and enzyme competition, and cost is normalised to the real price per effective daily dose, then benchmarked against the market. A verdict is KEEP, REPLACE, or DROP, and every claim carries its citation.

The demo's NMN verdict is the shape of it. The compound reliably raises NAD+, which is a surrogate marker, and a decade of human trials shows no consistent clinical payoff, at roughly seventeen times the category cost of an alternative with replicated effects. Tier D. Replace.

What runs today: the full evaluation experience, walkable end to end on sample data, built on the Claude API with a structured retrieval pipeline over named, checkable sources. Next: the live data pipeline and authentication on Supabase, and the repair layer, implemented behind feature gates and deliberately inactive in the demo, which performs only bounded, provenance‑grounded corrections before a verdict is allowed to stand.

Money cannot reach a grade. The rating side is built with no access to commercial data, eligibility flows one way from grade to listing and never back, nothing buys entry or position, and a product that loses its grade is removed regardless of the revenue attached to it.

AI trip planning is saturated, and nearly all of it draws on the same pool: aggregated reviews, sponsored listings, SEO content. Destinations converge on one itinerary, and a genuinely good local place loses to a larger one with more reviews. The agencies who know their country hold the knowledge that would fix this, and spend their days retyping it one email at a time.

LoreTrip inverts the priority. The agency curates the partners and places it will vouch for, with priority levels, notes, and seasonal relevance, and that layer becomes what the planner reaches for first. Travellers can still extend and swap, but every alternative comes from the same vouched pool, so customisation never drifts into generic. Where the agency has curated nothing, the planner falls back to named public sources and continues rather than stalling. The advisor can lock any element, and nothing reaches a client unsigned.

The reason to ask a local where to eat is not that they hold more data. It is that they hold judgment, and an aggregate score is precisely what destroys judgment. LoreTrip is an attempt to carry judgment through automation instead of averaging it away.

Three layers work today: the agency workspace, the agency‑controlled partner layer, and the platform layer handling agency billing. The reference build runs as a fully interactive experience, with a branded assistant, animated maps, a packaged itinerary screen, and a generated PDF trip guide, and the product is registered in the Anthropic Claude Partner Network. Current work is the white‑label tenancy itself, done slowly on purpose: tenant isolation, EU‑region hosting, and GDPR data‑processing terms have to be right before a second agency exists, let alone a tenth. Mijn Slovenië is the deliberate first tenant, monitored for usage, cost, and feedback, to stress the multi‑agency layer against one real business before expanding.

Povio originally built the site; the more interesting part is how the business runs today. Mijn Slovenië operates as an AI‑orchestrated business: an AI operating system handles the recurring work: social content, first‑line customer questions, trip planning and destination research, competitor tracking, and the back office, all inside a defined brand voice. My partner and I direct and review rather than do each step by hand, and nothing reaches a customer without a human approving it. I co‑run it as Founder & Creative Director.

It is also where I learned orchestration properly. The hard part of a system like this is not prompting; it is architecture: a hierarchy of instruction files that behave like departments, each scoped narrowly enough to stay coherent, so context does not bloat and the model does not drift into invention. I run it myself rather than delegating it, because that hierarchy is the difference between a chat that impresses once and an ecosystem that still works in six months.

AI earns its place when it keeps a real operation moving and extends a person's judgment, not when it only demos. The business is the test: instead of claiming AI can run operations, we run one on it, and keep a human in charge of what actually reaches a customer.

It started with paint: dilution, gravity, control. It changed the moment I stopped trying to reproduce a master and started listening to what the material was doing. V Á K U U M records real physical processes as they bleed, dry, and settle, then treats that footage as compositional material. Not scientific documentation; the same attention to how matter behaves that governs lab work, turned toward what it looks like up close.

The work moves in three stages, and the movement is the point. Acrylic on physical tiles first, poured and flooded and left to decide its own final form, because drying is only a verdict, not an ending. Then digital re‑readings of the same source. Then the audiovisual series, which gives the motion back the time a still image takes away. Made over several years, each piece records a process that happens once and cannot be repeated, and more than 100 works across the three series now share one visual language.

It also asked, early, the question the rest of the field is asking now: if a work is shaped with something autonomous, is it still human? I think yes. A person steering the machine has not stopped travelling; they may simply reach places they could not have walked to alone.