Dealbot

Find and buy second hand items with ease. Dealbot is an AI agent which can search for items over multiple marketplaces, identify fake goods, and negotiate prices.

Comment

DealBot

Inspiration

We've all been burned by sketchy online marketplaces. Whether it's overpriced items, scam sellers, or just poor deals, shopping secondhand can be risky and time-consuming. We wanted to build an AI agent that could do the heavy lifting - searching, evaluating, negotiating, and even purchasing, so users get the best deals safely.

What it does

DealBot is an autonomous AI shopping agent that:

  • Searches multiple marketplace listings based on your product query
  • Evaluates deals using intelligent scoring (price, condition, seller reputation)
  • Assesses scam risk by analysing seller behaviour, payment methods, and red flags
  • Negotiates with sellers via automated messages to get better prices
  • Recommends purchases based on value and safety
  • Executes transactions using an integrated wallet system
  • You describe what you want to buy, and DealBot handles the rest. Think of it as having a personal shopping assistant that never sleeps.

How we built it

Frontend (React + TypeScript)

  • Built a clean dashboard with real-time chat interface to track agent runs
  • Integrated Supabase for persistent storage of chats, products, negotiations, and transactions
  • Used TanStack Query for efficient data fetching and caching
  • Backend (Python + FastAPI)

Multi-stage orchestration system:

  • Stage 1 (Agent): Coordinates search, evaluation, negotiation, and recommendation
  • Marketplace Adapter: Abstracts marketplace APIs (Vinted, eBay, etc.)
  • Fake Marketplace: Test environment with seed data for development
  • LLM integration (Claude, Gemini, or local LM Studio) for intelligent decision-making
  • Scam detection using rule-based scoring on seller metrics and transaction patterns

Automation (Selenium)

  • Browser automation for direct marketplace integration (Vinted messaging, item interactions)
  • Handles login, cookie consent, location selection, and automated messaging

Infrastructure

  • Supabase (PostgreSQL + Auth + Edge Functions) for data and user management
  • Stripe integration for wallet top-ups and payment processing
  • Docker-ready for easy deployment

Challenges we ran into

  • RLS Policies & Database Design: Got caught by Row-Level Security policies that were preventing inserts. Fixed by removing buggy default values and creating proper permission rules.

  • Marketplace Scraping: Each marketplace has different DOM structures and anti-bot measures. Had to implement stealth mode in Selenium and handle dynamic popups (location selection, cookie consent).

  • Agent Consistency: Getting the LLM to follow structured outputs and make consistent decisions required careful prompt engineering and validation.

  • Real-time Sync: Keeping the frontend in sync with agent runs happening in the background was tricky. Solved with localStorage linking and automatic database syncing on completion.

  • Fake Marketplace Data: Initial seed data only had electronics. Had to regenerate with diverse product listings for proper testing.

Accomplishments that we're proud of

  • End-to-end autonomous shopping pipeline that actually works, from search to negotiation
  • Scam detection system that catches red flags humans might miss
  • Real-time agent run tracking with persistent storage of all decisions and messages
  • Multi-marketplace abstraction layer that makes adding new platforms simple
  • Wallet system with Stripe integration for autonomous purchasing
  • Clean, intuitive UI that demystifies what the AI is doing at each step
  • Fully testable with fake marketplace so you can iterate without hitting rate limits

What we learned

  • LLM agents need structure: Unstructured outputs from LLMs lead to chaos. We learned to enforce schemas and validation.
  • RLS is powerful but tricky: Supabase's Row-Level Security is excellent for multi-tenant apps but needs careful planning upfront.
  • Browser automation is fragile: Every marketplace change breaks selectors. We had to build resilient fallbacks.
  • Users want transparency: Showing the agent's reasoning, products considered, and negotiations attempted builds trust.
  • Autonomous purchasing needs guardrails: Rate limits, budget caps, and confidence thresholds are essential before letting an agent spend money.

What's next for DealBot

  • Live marketplace integration: Deploy Selenium bots to handle real Vinted, eBay, and Facebook Marketplace listings
  • Multi-marketplace arbitrage: Find price differences across platforms and automatically flip items for profit
  • Advanced negotiation: Train specialised LLM models on successful negotiation patterns from past runs
  • Seller reputation learning: Build historical profiles of sellers to improve scam detection over time
  • Mobile app: Native iOS/Android so users can monitor their agent on the go
  • Autonomous reselling pipeline: Buy underpriced items, relist them at market rate, and manage inventory
  • API for third-party integrations: Let other apps hook into DealBot's intelligence
  • Community marketplace: Share deals, negotiation strategies, and scam reports with other DealBot users

Built With

Share this project:

Updates