Misfits AI · Pre-seed
Twenty months from €1,000,000 in the account to one provable milestone: a demonstrably accurate sizing engine running on a live, fit-data-generating user base. Below the plan, the cost model behind keeping the app free — and the dataset that buys.
Three tracks from close. The sizing engine is the spine — it is the milestone the round is judged against. The other two exist to feed it and to prove it is worth something to a brand. One hire this round.
Accuracy is measured against real outcomes — sizes kept versus returned by live users — not offline benchmarks. The ad is written and ready to post at close, distributed through women-in-ML networks. Months 1–3 for a start date is the honest range for a senior applied hire, not a stretch target. Sources 3 and 4 — brand account linking and returns data — need brand deals and sit outside this round by design.
The cutoff is the point of it: polish is iterated forever elsewhere, and here it stops at month three so the engine gets the calendar time it actually needs. The number tracked on this track is the consented fit-data corpus, not installs — a vanity denominator for a data business. Creator spend stays only behind creators that clear payback. The web app ships in month four so a creator link opens the product directly instead of routing through a TestFlight install — it exists to serve corpus growth.
One progression, not three tracks. The stress-tests come three to four months after the ML engineer starts and are engineering validation first: mid-market brands invited to test recommendation-engine reliability against their live catalogs. What they return is what opens the brand conversations, and those are what convert into LOIs and pilot agreements. Founder-led throughout — there is no sales hire in this plan. No commerce-live date is committed here: the B2B fit-intelligence layer is what the seed round productizes, and promising its launch inside a pre-seed would be selling the next round's work as this one's.
Month 20 · the sentence the seed round opens with
“A sizing engine proven accurate on a live base of 2,500 users generating 25,000 fit-data records per month.”
Acceleration case · not plan of record
If the sizing-engine milestone lands ahead of schedule and the buffer allows: a sales hire is pulled into the pre-seed phase and brand pilots accelerate against it. This is the only circumstance in which a second hire appears, and it is contingent on a milestone that has not happened yet. The base plan above assumes it does not.
The app is free through this phase, and the question that matters is whether the round can carry that. It can: the whole free tier is sized to sit inside the 9% compute allocation. Start from a scenario, then take any input apart — the number to watch is the headroom against —.
Compute at month 20
How the spend is controlled
Generation allowance is behaviour-based. Higher caps are earned through purchases and kept items, so compute spend concentrates on the users who generate revenue and the densest fit data. The renders-per-user input above is therefore a blended average across the base, not a flat entitlement everyone receives.
The unilateral fallback: it exists, it is priced, and it can be switched on without anyone else's agreement. These are its numbers, kept visible rather than claimed.
The floor, if switched on
There is no CAC assumption on this page, for the same reason there is none on the allocation model: acquisition efficiency is the untested hypothesis this round exists to retire. The user base is set by hand and the plan reads what carrying it costs. Users ramp linearly from zero at close to the month-20 figure you set.
The intended model is commission-funded free usage, and it is not modelled here. Matching a generated look to a buyable SKU and earning on the sale is Phase 2 work, gated to the seed round and dependent on brand deals that do not exist yet. Projecting revenue from them on this page would be selling the next round's work as this one's, so the plan above is costed as if that revenue is zero — which is also why the headline figure is what free usage costs against its own budget line, not what it earns.
The subscription floor, held in reserve, does not cover the free tier — and is not meant to. A paying user is margin-positive by construction, so it covers paying-user inference several times over; the conversion at which it would cover the whole base is stated in the block so the gap is never implied away. Opening that block also applies the pricing ladder's five-generation free allowance, which is why compute steps down when it is open.
Compute is measured against the — allocation — — a month, — over the round, matching the Ask slide's — split at — average monthly burn. The Plan preset reconciles exactly: — users × — renders × — blended is —, plus — infrastructure base, is — — the allocation line, hit exactly at month —.
Cost per render. The Plan preset uses —, the blended rate the allocation model defaults to because production traffic mixes tiers (Flash —, Blended —, Pro —). The brief's marginal figure of €0.05 and the bottom-up model's — sit inside the slider's range; drag it and see how little the conclusion moves.
Consent rate defaults to —, in line with the observed per-generation rate (as of 2026-08-27: 26 of 56 external renders carry improvement consent; 2 of 6 activated testers). A six-person cohort is noise in either direction — the default rounds the generation-level rate, nothing more.
The subscription is illustrative, not locked. Pro at — is the headline tier; the mix toggle blends it with Basic — and Pro+ —. Roughly — is kept after Apple's 15% small-business rate and RevenueCat. Nothing on this page is netted against burn — the plan is costed as if revenue is zero throughout, because it is.