TurnDeal team receiving first place at the Sea × OpenAI Regional Codex Hackathon Taiwan
TurnDeal team developing the prototype at the Sea × OpenAI Regional Codex Hackathon Taiwan
TurnDeal team presenting the prototype at the Sea × OpenAI Regional Codex Hackathon Taiwan
Demo
Problem
Shopping AI Stops at the Published Listing
AI is becoming a shopping entry point, while marketplaces remain the place where shoppers verify information and complete purchases. Current assistants help with discovery, comparison, and checkout, but they still optimize among published listings. They do not form a buyer-specific counter-offer that can trade price against delivery, bundles, or relationship benefits.
Answers questions, recommends products, and tracks prices, but it does not form a counter-offer for one buyer.
Amazon Sparky
Supports conversational discovery and basket building, but it does not negotiate across competing sellers on the buyer's behalf.
ACP / UCP
Connects catalogs, checkout, and payments, but does not define how buyer and seller agents form a validated deal before purchase.
Proposed Solution
A Platform-Owned E-commerce Negotiation Layer
01
Platform Data Layer
Live product, inventory, rating, logistics, return, and member-benefit data stay with the marketplace. The Formatter turns a natural-language need into hard constraints and ranked preferences for discovery.
02
Agent Negotiation Layer
Private buyer branches negotiate with eligible seller agents in parallel. Each branch protects the seller's policy and context, while synchronized rounds share only anonymized, backend-validated market terms.
03
Trust and Validation Layer
Sellers configure 22 offer constraints. The backend checks ownership, inventory, total price, delivery, add-ons, and policy limits before publishing an immutable offer for the Evaluator to rank.
04
Decision and Checkout Layer
The React interface shows ranked Deal Cards on desktop and mobile. A shopper explicitly confirms one offer, which the platform revalidates before opening the simulated ACP checkout flow.
System Architecture
How TurnDeal Forms a Trusted Offer
Private buyer branches negotiate in parallel before the platform validates, ranks, and checks out a selected offer.
Step 01
Buyer Request
A shopper describes budget, delivery deadline, and product preferences in natural language.
Step 02
Formatter
Converts the request into verifiable hard constraints and ranked preferences.
Step 03
Orchestrator
Selects eligible sellers, freezes the request context, and creates isolated negotiation branches.
Step 04
Private negotiation sessions
Each seller sees only its own policy, product context, and branch history. Synchronized rounds can share anonymized, backend-validated market terms.
Seller-defined price strategy
Buyer Branch ASeller A
Request for Quote (RFQ): buyer needs and validated market terms
Seller reply: accepted, counter-offer, or declined
Own floor price and delivery terms
Seller-defined speed strategy
Buyer Branch BSeller B
Request for Quote (RFQ): buyer needs and validated market terms
Seller reply: accepted, counter-offer, or declined
Own logistics and service terms
Seller-defined bundle strategy
Buyer Branch CSeller C
Request for Quote (RFQ): buyer needs and validated market terms
Seller reply: accepted, counter-offer, or declined
Own add-on and benefit terms
After backend validation, anonymized market terms feed the next negotiation round
Step 05
Offer Validator
Checks seller ownership, stock, total price, delivery, add-ons, and policy limits before publishing an immutable offer.
Step 06
Evaluator
Ranks the complete eligible offer set without seller-private policy data.
Step 07
Shopper Approval
The shopper reviews Deal Cards and explicitly confirms a selected offer.
Step 08
ACP Test Checkout
Revalidates the accepted offer, then creates a simulated ACP checkout and payment flow.
Implementation & Outcomes
A Controlled Multi-Agent E-commerce Workflow
Multi-Agent Pipeline
We designed the pipeline that connects request formatting, catalog discovery, private negotiation, offer validation, evaluation, and checkout confirmation. Shared contracts keep the frontend, runtime, and tests on one negotiation model.
Parallel Negotiation
We implemented five seller branches that negotiate in parallel with synchronized rounds. The backend commits a shared revision only after it validates the prior round, keeping competitive information useful without exposing seller-private terms.
Seller Constraints & Validation
We modeled 22 commercial constraints for seller configuration and implemented offer validation, so the system evaluates only offers that meet the seller's inventory, price, delivery, add-on, and policy boundaries.
Responsive React Experience
We built the React interface so the buyer can enter requirements, follow progress, review Deal Cards, and confirm a choice through one responsive flow across desktop and mobile devices.
Sea × OpenAI Regional Codex Hackathon Taiwan
Champion
Built by a four-person team in one day
Private Negotiation Branches
5 Sellers
Each branch keeps its policy and context isolated
Seller Commercial Constraints
22
Guardrails enforced before an offer becomes eligible
Limitations
What Remains to Be Validated
TurnDeal is a hackathon prototype. Product information and prices are scraped from e-commerce sites, and the current database covers mice only. Seller floor prices come from LLM-generated values based on predefined personas, not live merchant policies. Production use would require verified catalog and pricing feeds, real seller integration, stronger policy and fraud controls, privacy review, payment compliance, and experiments with real shoppers and fulfillment outcomes.
