DoorDash_Gordon

GORDÓN turns every DoorDash merchant inventory scan into promote, hold, or auto-disable—using Zero weather/safety, Nexla ops data, OpenAI copy, and Pomerium role gates to cut waste and stockouts.

Comment

Inspiration

I run an If I were a PM series where I turn product ideas into real MVPs — check it out here: https://www.youtube.com/@BuildWithAshna. For this hackathon, the DoorDash merchant problem was obvious: restaurants already lose money on spoilage and stockouts, but inventory still sits in a fridge while DoorDash demand, weather, and food-safety news move in real time. GORDÓN is the PM bet that every barcode scan should become a decision — promote, hold, or auto-disable — not just a log entry.

What it does

GORDÓN is a DoorDash merchant inventory co-pilot. Staff scan (or enter) a barcode; the app: Looks up the product and links/creates a recipe Pulls live weather + foot-traffic signals and food-safety/outbreak news Reads real ops context (orders, waste, prefs, inventory) Decides promote, hold, or disable after the next order based on qty + demand math Lets only Manager/Owner approve promos or change availability Math stays deterministic; language models only write promo copy and explanations.

How we built it Next.js + Prisma/SQLite for the merchant app and transactional inventory Zero.xyz for barcode, weather, and food-safety news Nexla MCP → Neon Postgres for operational reads that ground recommendations OpenAI for promo copy only (never numbers) Pomerium-style JWT identity for group → role gates (STAFF can’t approve; Manager/Owner can) Deterministic engines for excess inventory, weather-adjusted DoorDash + walk-in demand, and smart availability

Challenges we ran into

Keeping sponsor boundaries clear: Nexla reads ops data; inventory writes go to Neon so the MCP always sees fresh scans Making any new SKU (not just Pepsi demos) create a recipe and get a smart promote/hold/disable decision Making roles actually demoable without a full IdP — signed Pomerium-shaped JWTs + UI that blocks STAFF from Accept Not letting the LLM invent financials — OpenAI phrases; engines calculate

Accomplishments that we're proud of

End-to-end scan → recipe → weather/ops/safety → action card in one intake flow Smart decisions for any scanned product, not hard-coded demos Live Nexla MCP reads + Neon write-back so inventory isn’t fake sample data Role-gated approvals that judges can flip live (STAFF locked → Manager unlocks) A compact Zero food-safety warning on the dashboard that still tells the outbreak story

What we learned

Merchant tools win when the first action is obvious after a scan Ops data + weather beat generic “AI recommendations” for waste Sponsor integrations need a honest fallback (demo JWT, mock Zero) that still proves the architecture Separating math vs language keeps the product trustworthy

What's next for DoorDash_Gordon

Live Pomerium proxy + IdP groups in production Real DoorDash menu/promo APIs beyond simulation Multi-location merchant chains and supplier lot tracking for outbreak pauses Closed-loop learning from accepted vs declined promos via Nexla prefs Deeper Zero x402 paid data paths and richer shelf-life models

Built With

  • cursor
  • nexla
  • openai
  • pomerium
  • zero
Share this project:

Updates