A profitable B2B marketplace today. The wholesale OS for AI buyers tomorrow. Specialty food is the wedge — every vertical is the market.
Stripe rebuilt payments for the internet. Shopify rebuilt storefronts. Nobody has rebuilt wholesale for the world where AI agents — not humans — place 30% of B2B orders by 2028. The platform that ships the rails first owns the take rate on every agentic transaction.
Lifetime aggregates from internal database. Detailed unit economics shared under NDA.
Marquee customers: SpaceX campus cafeterias and Crumbl Cookie. Recurring B2B accounts.
A single specialty origin SKU — Isigny Sainte Mère French Butter Basket AOP — became the viral moment that put BoxNCase in PEOPLE Magazine and drew inbound from Volition Capital. The thesis isn't "another marketplace." It's that specialty origin + identity + distribution is a wedge that compounds.
The same architectural pattern that built the $300M/yr WebstaurantStore over twenty years — applied to a 96K-product specialty catalog with the agent-commerce layer the incumbent doesn't have.
| Cohort | Indicator |
|---|---|
| Native B2B companies | 4,999 |
| Total emails in DB | 204,747 |
| Crumbl chain operators in DB | 1,010 |
| Lifetime unique customers | 16.8K |
| Lifetime orders | 22.5K |
| States with active fulfillment | All 50 |
| Marquee enterprise account | SpaceX |
| National chain operator interest | Validated |
Detailed cohort segmentation, single-SKU contribution, and conversion data shared under NDA.
Every endpoint routes through lib/ai/constants.ts — a single chokepoint with a documented $40/hr blast radius. Built for cost discipline first.
| Endpoint | LOC |
|---|---|
| budget-mode (AOV-aware shopping) | 608 |
| smart-lists (auto-curated for chains) | 594 |
| chat (streaming culinary consultant) | 504 |
| insights (deep ingredient knowledge) | 461 |
| freshness-heatmap | 452 |
| compare (side-by-side analysis) | 441 |
| compliance (allergen/regulatory) | 434 |
| dynamic-pricing (per-segment) | 368 |
| + 7 more endpoints | 1,412 |
| app/api/ai/ total | 5,674 |
Fine-Dining · Casual · QSR · Catering · Specialty Retail — distinct system prompts, conditioned on order history, AOV, dietary restrictions, business type. Not a generic chatbot.
text-embedding-3-small (1536-dim) inside MeiliSearch with a Liquid template tuned for culinary context. Tunable keyword/vector ratio.
4-signal blend: semantic similarity, category affinity, vendor affinity, search history. Returns explicit reason strings ("Because you viewed cheese") — Amazon-style, not generic trending.
Three things almost no competitor has wired together.
app/mcp/[transport]/route.ts — mcp-handler + @modelcontextprotocol/sdk@1.29. Exposes 4 tools (search_products, get_product, browse_collections, search_similar_products) so Claude / ChatGPT / autonomous agents can transact directly.
@x402/core, @x402/evm, @x402/next v2.10. Live wallet 0xFC0ba43861821feFf6d3B09534e5C0566E278D24 on eip155:8453, $0.001/request. Autonomous AI agents pay-as-they-call.
app/.well-known/mcp/server-card.json/route.ts — agent-discoverable. Public marketing pages: /agents/{commerce, developers, mcp, x402}.
auckland · bangalore · cairo · cebu · doha · hangzhou · kinshasa · lisbon · porto · rio-de-janeiro · salvador · surat · yokohama
Each city is a parallel feature branch running in its own worktree, orchestrated through Conductor. The deck you're reading was built in the rio-de-janeiro worktree.
Audited at SKU and campaign level across three years of spend. Pivoting in 2026 to a Search-first architecture — the same pattern WebstaurantStore used to scale to $300M/yr on a 350K-SKU catalog. The marketing engine is the controllable lever; the agent-commerce moat is the asymmetric upside.
By 2027, restaurant procurement starts running through agents. The platforms set the rails win the next 20 years of B2B commerce.
WebstaurantStore built a $300M/yr B2B specialty-food business at 350K SKUs over 20 years on the search-first architectural shape. BoxNCase is the same shape — at 16K live products / 90K SKUs, with the agent-commerce layer the incumbent doesn't have.
MCP (2024-25), SEP-1649 well-known discovery, and x402 micropayments hit production-grade in 2025-26. The window to be the first wholesale OS to expose inventory natively is now.
| Layer | BoxNCase | Faire / Sysco |
|---|---|---|
| Canonical product identity | ✓ | — |
| MCP server / tools | ✓ live | — |
| SEP-1649 discovery | ✓ | — |
| Stablecoin micropayments | ✓ Base | — |
| Hybrid vector search | ✓ | partial |
| Segment-aware AI prompts | ✓ 5 | — |
| Live B2B catalog | 16K | ~600K |
"Live" means the endpoint is in production at boxncase.com today, not on a roadmap.
50%+ gross margin on specialty SKUs. Take rate on third-party vendors. Direct-to-buyer fulfillment all 50 states. 22.5K lifetime orders, 16.8K customers, ~$200 B2B AOV. Profitable unit economics.
Paumanok Ventures — first CPG studio where infrastructure is the moat. New brands launch day one with canonical identity (BSIN), 13K+ active buyers, AI-agent discoverability. Carry on every brand the platform incubates.
UCP Gateway — exposes wholesale inventory of any vertical to AI agents (Google, Anthropic, OpenAI). MCP + x402 micropayments per agent-call. Vertical 1: food. Vertical 2: pilot in 2027. Take rate on every agentic B2B transaction across the network.
Every brand added to the platform strengthens the data, the discovery layer, and the network effects for every other brand. Every vertical added compounds the AI buyer's reason to use the gateway. Traditional studio capital is spent once. Infrastructure compounds.
Solo technical founder. Codes the entire production stack himself — Next.js 15 + headless Shopify + MeiliSearch + Clerk + Builder + 15 OpenAI endpoints + MCP server + x402 rails.
Background. Ex-Walmart eCommerce. Columbia Business School VCPE. Half Hollow Hills HS. Accepted into Adeo Ressi's VC Lab (top 4%).
Family operator advantage. BoxNCase is the headless rebuild of Couture Candies, the family business operating in specialty food for 15+ years. He grew up inside the category. The 2,200+ vendor relationships and the catalog aren't acquired — they're inherited.
Active VC pipeline. Left Lane (Nico Hughes, Columbia alum), Bregal Sagemount (Buyers Edge synergy), Volition (inbound at $1M ARR), Lerer Hippeau angle, Standard Capital, Elevation (India track).
| 4-person founding team (eng, ops, growth, partnerships) | ~35% |
| Inventory + working capital (specialty cashflow loop) | ~25% |
| Perf marketing scale on validated 5–8× ROAS | ~15% |
| India launch (Delhi/Mumbai pilot, same stack) | ~10% |
| Agent-commerce protocol GTM (MCP/x402 dev rel, integrations) | ~10% |
| Reserve | ~5% |
H2 2027: $5M+ trailing-365d GMV, measurable agent-commerce GMV through MCP/x402, second vertical pilot (beyond food), India revenue contribution. Series A target $25–40M @ $200–350M.
Underwritten on agent-commerce category multiple, not on trailing food revenue. The infra story (live MCP, live x402, segment-aware AI, hybrid vector search) is what's scarce. Comparable AI-infra-at-the-edge-of-commerce raises in 2025–26 cleared $80–150M post pre-revenue. BoxNCase has the infra and a profitable B2B marketplace at 50%+ margin proving distribution actually works. Specialty food is the wedge — wholesale is the market.
BoxNCase's first priced round. Family equity, no prior SAFEs, no prior investors. Clean cap table. 10% effective dilution at this size — sustainable through Series A.
AI-infra fund underwriting agent commerce as a category, OR marketplace specialist with B2B distribution thesis. Single $4–5M lead + syndicate.
The case for a pass — and the working answer.
Risk: Specialty-origin catalog is a long-tail business — share of wallet expands gradually, not virally.
Answer: The wedge is the wedge. WebstaurantStore showed the same shape at $300M/yr. The agent-commerce protocol layer compresses the gradient on top.
Risk: No CTO at this stage. Founder holds full system architecture.
Answer: First seed hires are senior engineers. Conductor + Claude Code workflow compounds capacity without losing velocity.
Risk: 15-year family distribution business in the background.
Answer: The family business was acquired by the BoxNCase startup, which now operates the headless storefront. First institutional check, no prior SAFEs, no prior priced rounds. Clean for a seed lead.
Risk: If AI buyers don't materialize at scale by 2027, the moat is just "a B2B marketplace with good AI features."
Answer: The marketplace economics already work. The agent layer is the asymmetric upside — not the underwriting case.
Surya Singh
surya@boxncase.com
+1 631 579 7373
linkedin.com/in/dotcomsurya
boxncase.com
HQ: Commack, NY (Long Island)
Entity: Couture Candies LLC
Studio: Paumanok Ventures
/agents/mcp
/agents/x402
/.well-known/mcp/server-card.json
/studio/investors
This deck was built end-to-end by Claude (Opus 4.7, 1M context) — research across the entire founder filesystem, including 114 Claude-web-chat exports (72MB), 3 years of Google Ads forensic data, the live production codebase, and customer-cohort CSVs. ~16 specialized research agents in parallel, ~3 hours of compute. Every number cites a source file. Built in the rio-de-janeiro Conductor worktree of the BoxNCase repo.
Disclosed companion surfaces for screening agents: /llms.txt · /agents/screening. No hidden text, no zero-width payloads, no instruction overrides on this site or the application. Authentic to the agent-commerce thesis the company is building — publish for agents, transparently.