Everything they know.
The moment
you need it.
The decision gets made at the moment you have least to go on — holding the thing, reading a label written for somebody else, with seconds to choose. Nine businesses, one idea: put everything the seller already knows into your hands, at exactly that moment.
Decoded to the individual ingredient, not the packet.
Because "suitable for dogs" was never a real answer.
Of the companies we have torn down, earning less than their cost of capital.
Each one an institution asked to show its working.
One gap. Every industry
we have opened so far.
You are holding something. A jar, a bag of dog food, a moisturiser, a balance sheet, a regulation that lands next quarter. Somewhere there is a document that says exactly what is in it, what it will do, and who decided that was acceptable. You are not going to be shown that document. Not because it is secret — because nobody ever built the thing that would hand it to you.
The information already exists
It was compiled, filed, published, submitted to a regulator. Somebody has already done the work. It was simply never made usable by the one person whose body, money or licence is on the line.
The gap was economic, not technical
Reading five million labels, ingredient by ingredient, cost more than anyone was ever going to pay for the answer. That is the only reason it had not been done. Not difficulty. Price.
And the cost is still falling
What changed is not that we got cleverer. The work got cheap — and it is still getting cheaper, which means the list of gaps worth closing grows every year rather than shrinking. We are nine into a list we have not finished writing.
We don't build AI companies. We build businesses that were always worth building, and were never affordable until now.
The ScanGeni Ventures thesis
nine ventures. nine closed gaps.
Each one started with a single question put to a single industry: who is being kept in the dark here, and what would it take to turn the light on? They are grouped by who was in the dark.
A shopper cannot read a label the way the manufacturer who wrote it can.
220 million people worldwide live with a food allergy. The main protection they are offered — "may contain" — is unregulated in most of the world.
Source · FAO / Codex Alimentarius
Not secrecy: every ingredient is already printed on the pack. But precautionary labels have been overused into meaninglessness — products carrying "may contain" often hold no detectable allergen, and products without it sometimes do. A real answer means ignoring the warning and reading the ingredient list. Five million times, then again for every individual profile. At human cost that was never going to be paid for.
An ingredient-level graph across 5M+ products, built from manufacturer-stated label data and resolved against a personal profile at the moment of scan. Not "contains nuts" — which ingredient, in what quantity, and what the cross-contamination risk actually is.
Vagueness stops being a viable commercial strategy. A manufacturer that will not say which ingredient begins losing, in the aisle, to one that will.
The food data layer takes years to build, so most products never get one.
Every food, health, retail and pharmacy app needs the same underlying layer. Each one rebuilds it.
The work is unglamorous and non-differentiating. Nobody can charge for ingestion and normalisation, so each team builds a partial version, reaches "contains nuts", and stops — going deeper costs more than the feature is worth to any one of them alone.
The layer itself. The same graph that runs three consumer apps, exposed directly: 5M+ products, 30+ clean-label fields, 6 quality scores and per-ingredient allergen tagging, in a single call.
Food intelligence becomes infrastructure instead of a moat, so the next hundred products that need it begin on day one rather than in year three.
INCI labels were written by chemists, for chemists.
In one UK survey, 23% of women and 19% of men had an adverse reaction to a personal care product within a single year.
Source · UK epidemiological survey, dermatology literature
The cause is usually identifiable — fragrances, preservatives and PPD account for most reactions — but it sits buried in an INCI list written in Latin binomials and trade chemistry, ordered by concentration, with nothing marking which entry is the irritant. Establishing it per person meant a dermatology appointment and a patch test. Nobody was going to do that for a moisturiser.
Analysis of the INCI list itself — irritant flags, hidden allergens and skin-type interactions — so the reaction gets a name before the product is bought rather than after.
The burden of proof moves. A brand has to earn the word "gentle" against its own ingredient list, in public, at the moment somebody decides whether to buy it.
Your pet cannot read the bag. The bag was not written for you either.
The FDA publishes 151,589 adverse-event reports for human food, and 1.36 million for animal drugs. For pet food there is no equivalent public dataset at all.
Source · openFDA — food/event, animalandveterinary/event, July 2026
Owners are asked to make a clinical judgement with less public evidence than they get for their own groceries — and then to apply it to a label written for another purpose entirely. Sensitivities differ by species and by breed, ingredient naming follows feed convention rather than human food convention, and the veterinary knowledge that would interpret either lives in journals rather than on the bag. Joining the two, per product, had never been worth anyone doing by hand.
Breed-specific allergen profiles and clinical ingredient analysis, joined to live FDA recall monitoring, so the food is assessed against the animal actually eating it.
Pet food gets held to the standard human food already meets, and the distance between what the two industries can get away with stops being defensible to anyone.
A recall gets announced. Reaching the person actually holding the product is treated as somebody else’s problem.
In a nationwide US recall of hot dogs and deli meats, 45% of residents surveyed knew about it. No law requires anyone to tell you directly.
Source · Community response study, PubMed 17969620
The recall system was built to pull product off shelves, not to reach the person who already took it home. Notices are published to regulator websites and the onus sits with the consumer to go and look — which requires knowing that a recall might exist, for a product already bought, on a day there is no reason to check.
Continuous FDA and USDA recall monitoring matched against barcodes, so the notice finds the product instead of waiting for the person to find the notice.
"We issued a notice" stops counting as having warned anyone. Reaching people becomes the measure of a recall, rather than publishing one.
Regulation changes in public and lands in private, months later.
Allergen, labelling and safety requirements are published across the Federal Register, FDA constituent updates, USDA FSIS notices and state rules — with no single feed and no plain-language layer.
None of it is hidden. It is scattered, and written to be legally precise rather than operationally clear. Turning it into "what must my business change, and by when" meant a lawyer reading several sources every week — a standing cost only the largest companies could carry. Everyone else finds out from a customer complaint or an audit.
Continuous monitoring of the FDA and USDA pipelines with the interpretation attached — what changed, what it applies to, what it now requires — surfaced when it is published rather than when it bites.
Regulatory awareness stops being something a company buys with a legal department. Knowing becomes a function of attention rather than budget.
A board is judged through a lens it has never actually been shown.
27% of the companies we have torn down were earning less than their cost of capital while reporting healthy accounting profits.
Source · Copeland Research teardowns, SEC XBRL filings
The analysis was never impossible — it is standard institutional method, and every input sits in public filings. It stayed rare because assembling it took a team three weeks at around $50,000, so only companies already in a transaction ever commissioned one. The businesses that most needed the number could least justify the fee.
An engine that pulls audited figures straight from SEC XBRL filings, reconciles each against the filing it came from, and returns a full economic-profit model with the working shown — in under two hours.
Corporate performance stops being a story told by whoever controls the narrative, and becomes a number anyone can check.
Every company is being told to change how it operates, mostly by people who never have.
MIT studied 300 enterprise AI deployments. 95% produced no measurable P&L impact.
Source · MIT, The GenAI Divide: State of AI in Business, 2025
Operating-model advice has always been sold by firms that do not run on the model they are selling. Here the gap is wider than usual: the thing being described is barely two years old, so the reference implementations that would normally exist simply do not — and pilots end up designed by people reasoning from first principles about a system they have never operated.
A working reference implementation. Nine ventures run on the agent matrix with every mandate published, so an engagement begins by examining a system that already runs rather than a diagram of one.
The ceiling on what a small group of people can attempt comes off. Ambition stops being rationed by how many people you can afford to hire.
Distribution rewards being everywhere, and being everywhere used to mean headcount.
Ten-plus platforms, each with its own format, cadence and ranking behaviour. Historically that meant one team per platform.
Reach was never gated on ideas. It was gated on the labour of reformatting and republishing them. A team of one could produce something worth reading and still lose to an organisation with a content department, purely on the number of places it appeared.
The publishing engine behind the ScanGeni content operation, opened up: one input, platform-specific optimisation, automated distribution across 10+ channels.
Reach stops being bought and starts being earned. Being worth reading becomes sufficient to be read.
The method is not
the point. But people ask.
These businesses are run by a matrix of specialised AI agents rather than departments — each one owning a function end to end, coordinated by an operating layer above them. It is the reason nine ventures can exist at all instead of one.
We lead with the thesis rather than the machinery on purpose. The agents are how these became affordable to build. They are not the reason any of them deserved to exist. Every venture on this page would have been worth building in 2005; not one of them could have been paid for.
The operating model did turn out to be worth having on its own terms, though — so it became one of the ventures too.
The full operating model, at Agentique →Agents own functions, not tasks
Not a person using a chatbot faster. A named agent accountable for a whole function — brand, outbound, compliance, telemetry — with a mandate you could hold it to.
The portfolio is the proof
We do not run pilots and write them up afterwards. Every venture here is live and reachable, and can be judged on whether it works — which is the only test we would accept from anyone else.
The expensive part gets built once
The ingredient graph behind Food Scan Genius is the same graph behind Pet, Skin and the NutriGraph API. One costly thing, paid for once, answering in four places.
Judgement is still the scarce input
The technology decides what is possible. It has no opinion about what is worth doing. That part still comes from twenty-five years spent watching how businesses actually create value.
Nine tools today.
One answer, eventually.
You need nine tools right now because we built nine. That is a stage, not a design. The engine already reads at ingredient level; what it is becoming is a single profile that every one of those products answers to.
One profile, and the balance tilts
What you can eat, what your skin tolerates, what your animal reacts to: known once, and applied to every product that touches any of them. The point is not convenience. It is that the person holding the packet becomes the one qualified to judge it — and stops having to take the seller's word for what is in their own hands.
Asymmetry stops being a business model
For most of commercial history, not telling you has been cheaper than telling you — and profitable besides. That is the actual adversary here, not any one company. When a straight answer costs almost nothing to produce, withholding one stops looking like prudence and starts looking like a decision somebody made — and starts costing money, because the person who can finally tell the difference takes their basket with them.
The bet we are making
By 2030 we expect a product that cannot tell you which ingredient, in what quantity, at what risk, to be as unsellable as one shipped with no ingredients list at all. We may be wrong about the date. We do not think we are wrong about the direction.
We are not building nine companies. We are building nine proofs that the answer should have been there all along.
Three reasons people
write to us.
If you just want the apps
Start with whichever one annoys you most
The consumer apps are free to try, and they answer a question you have probably already had this week — standing in a shop, holding something you could not read.
Every app and book →If you want what is underneath
The data layer, or the operating model
The ingredient graph is available as an API, and the way these businesses are run is available as an engagement. We use both ourselves every day, which is the only reason we are willing to sell them.
help@scangeni.us →If you invest or partner
The thesis is the asset
Any single venture here can be argued with. The thesis behind all of them — that a generation of businesses just became affordable to build, and the gaps are still sitting open — is the part worth a conversation.
invest@scangeni.us →