Taste Lab is a chat-first meal-planning app backed by an autonomous recipe-production engine — built by one person for ~$850. A human team would bill €2.5–3M to rebuild it.
A 12-agent audit of the live repository measured ~433 senior-engineer-weeks — about 8.3 person-years of shipped software: ~229 modules (212 already live), 56 database tables, 4 apps. Beyond the consumer app sit three internal platforms most pre-seed teams never build — Mise (recipe-generation engine), Ledger (behavioral analytics) and Taster (automated chat QA) — a third of the whole build.
Rebuild effort by subsystem · senior-engineer weeks · tap a bar
Every dish runs skeleton → enrichment → validation → persist behind safety, nutrition and realism gates, then a food-photo swarm shoots it — verified and styled, end to end. Modern, typed, test-covered stack: Next.js 15 · TypeScript strict · PostgreSQL · Drizzle · Zod · LangGraph · React Flow · Playwright.
Recipes and chefs are batch outputs, not handcrafted entries — long-running autonomous runs that produce them at volume, each fully verified and styled. Marginal cost ≈ $0; the only ceiling is run time.
The SKU importer + canonicalization engine snap raw store inventory onto one ingredient model — vendor-neutral, allergen-safe, deterministically sourceable.
Today · 4 stores · 20,725 SKUs → any regionEach dish is generated, validated against safety/nutrition/realism gates, then shot by a food-photo swarm. No human in the loop.
Tested capacity · 10,000 verified recipes & photos / dayReaching a million verified, styled recipes and thousands of chef personas is compute hours, not headcount or capital.
Today · 136 chefs · 309 batches · 1.06M stepsReal, published output — swipe the deck, then open the full detail
Point the engine at a store it has never seen and let it run. In a single autonomous day — no human in the loop — it absorbs the inventory, casts a kitchen, writes the menu, and shoots every plate.
The importer + canonicalization engine snap a store's raw inventory onto one canonical ingredient model — vendor-neutral, allergen-safe, deterministically sourceable.
Persona generation gives each chef a distinct voice, palate and ingredient bias; a portrait swarm shoots their likeness. A whole collective, from scratch.
Each dish runs skeleton → enrichment → validation behind safety, nutrition and realism gates — grounded in the very SKUs absorbed at hour zero.
A food-photo swarm styles and shoots each recipe — studio-grade, one image per dish. No set, no stylist, no shoot day.
A grocery the engine had never seen is now a verified, photographed catalogue — in one autonomous run. Marginal cost ≈ $0; the only ceiling is run time.
The whole build ran on flat-rate subscriptions — 3 months of Claude Code Max + 5 of MiniMax, ~$850. At metered API list-pricing the same usage (1.24B real tokens across 7 models, de-duplicated by request) would bill ~$58.5k — a ~69× gap. The asset itself is priced bottom-up from the code: a replacement-cost floor, cross-checked against two other valuation lenses.
Replacement cost, built bottom-up from 433 weeks of code
Raw labor = 433 weeks × €3,000/senior-week. The premium prices what a headcount can't buy instantly: coordination, paid-for dead-ends, slowly re-derived domain knowledge. Across a defensible rate band the floor stays inside €2.3–3.6M; central estimate ~€2.5–3.0M.
What a competitor actually pays vs. what it cost
433 weeks costed bottom-up + premium + data assets.
EU pre-seed comps, up for shipped IP, down for zero revenue.
Back-out from a ~€8–12M seed post-money at standard dilution.
Three independent methods converge on a €2–4M pre-revenue anchor. The ~$850 raises the efficiency case — it does not lower the floor a competitor still has to pay.
Taste Lab turns "feed my family this week" into a grocery basket against real SKUs. The basket is the monetizable unit: commission on referred GMV (typically 3–5%) plus a thin premium tier. Spain is the live beachhead — 19M households, a >€100B grocery market only 3–4% online, with Bonpreu's 20,725 SKUs already mapped. Drag the engine below.
Revenue/HH = basket × baskets/mo × 12 × take rate. Implied valuation applies the scenario's forward-ARR multiple (7× bear · 8× base · 10× bull), cross-checked against a 0.35–0.6× influenced-GMV multiple. A US household is worth ~1.6× the EU one ($110 basket, 2.5×/mo → ~$165/yr) but faces multiples-higher CAC; US lives in the bull case.
The base case is the engine run once for Spain, then ported country by country — Portugal, Italy, France — each reusing the same machine plus a local catalog. Below is the implied base-case valuation, quarter by quarter, with the two financing rounds that fund the ramp.
Start relationships + a small strategic angel round on shipped IP and the taste model. Run the non-dilutive grant track (EIT Food, Horizon CL6) in parallel.
Raise ~€1.5M once Spain's conversion loop, a second retailer and an evidenced take rate clear the gates. Funds national rollout. ($12–20M on a US cap table.)
Raise ~€8–12M after the Portugal cross-border proof. Funds Italy and France entry — the step that pushes the bull case past half a billion.
As of Dec 2025, the full meal-intent-to-checkout loop ships inside ChatGPT via Instacart, and rich chat UI is a documented Apps-SDK pattern — so neither is a differentiator. But incumbents do SKU-level inference, not a dish-level taste graph; genuine pantry-state lives only inside a $4,000 fridge; and no verified product combines dish-level taste + software pantry-state + WhatsApp + a canonical-catalog production stack into one assistant. That's the bet.
| Player | Loop → checkout | Canonical catalog | Software pantry-state | Dish-level taste graph | WhatsApp channel | Gen-UI in chat |
|---|---|---|---|---|---|---|
| Taste LabUS + Spain/EU | Planned | Yes | Partial | Shipped | Phase 1 | 10 intents |
| Instacart × ChatGPTUS | Yes | Yes | no | SKU-level | no | Yes |
| Kroger + GeminiUS | Rollout | Yes | no | SKU-level | no | some |
| Samsung Food + Fridgeglobal, split | redirect | no | Fridge HW | SKU-level | no | app UI |
| MercadonIASpain, unofficial | no | scraped | no | no | no | text |
Recreating just the published catalogue — 3,102 developed recipes, each styled and photographed — would cost ~$0.78M of human recipe-dev + food-styling + photography ($100 dev + $150 photo per dish), on top of the engineering floor. And it recurs at ~$0 marginal cost.
104 active chef personas (136 built) and a swipe-trained, dish-level affinity model — a proprietary preference dataset that compounds with every session and can't be bought, only earned. Plus Mise (catalog), Taster (eval loop), Ledger (analytics): a stack that lets a one-person team out-ship incumbents with 1000× the headcount.
The EU's food-waste cut is now legally binding, waste is concentrated in homes, and the dominant failure — over-buying — is exactly what planning + pantry-state attacks. Every corporate response so far is supply-side (markdowns, surplus apps). Taste Lab is the only one working the demand side — which turns a binding target into product positioning, non-dilutive capital, and a retailer wedge.
Horizon Europe CL6 (~€4M/project), EIT Food FAN, EIC Accelerator — a grant track that extends runway without dilution.
The WRI "$1 → $14" waste case doubles as the signed-retailer credibility unlock — the milestone that kills scraper-risk.
Cited honestly: app features alone show small average effects in the literature; the headline reductions came from intensive human coaching. The wedge is positioning, capital and a B2B door — not an efficacy claim.
A de-risked, eight-person-year build with a production stack most teams reach only after a Series A — shipped by one founder for ~$850. We pitch on evidence, not roadmap: proof moves the number, not scope or geography.
On shipped IP, the taste model and the production stack — while the gates below get cleared.
Run the institutional round once Spain's loop is proven. Funds national rollout across Mercadona / Carrefour / Dia.
After Portugal proves the playbook ports. Funds Italy and France — the venture-scale ramp.
This is a private look at what one founder has built. May we have your name?