FlipKit!

Upload a photo of any item into cash by using an to determine its market value. With one tap, our voice calls buyers to negotiate the highest offer on your behalf, eliminating the friction of selling.

Comment

Inspiration

Imagine you're at a garage sale. You spot an interesting item, but have no idea what it's worth. Or maybe you're cleaning out your attic and find a box of old trading cards, you wonder if you're sitting on a goldmine. But figuring out the next steps is hard. Finding the true market value, tracking down the right buyers, and dealing with awkward negotiations means wasted time and annoying haggling.

We wanted to build something that eliminates every bit of that friction, a platform where a single photo is all it takes to go from "what is this worth?" to "I just got paid."

What it does

FlipKit turns a single photo into cash in your pocket. Here's how it works:

  1. Snap a photo — Point your camera at any item and upload it.
  2. Instant identification — Our AI vision model (Claude Opus 4.6) identifies exactly what the item is, including brand, model, year, condition, and notable features.
  3. Pricing swarm — Five parallel Nemotron AI agents (orchestrated using the Claude SDK, with Opus 4.6 as the orchestrator) simultaneously scour eBay, Mercari, Amazon Resale, Facebook Marketplace, and Craigslist to determine the most accurate market value, factoring in condition, demand trends, and seasonal pricing.
  4. Local buyer matching — We search Google Maps to find pawn shops, specialty stores, and used goods buyers within 10 miles of your location, prioritized by who's most likely to pay top dollar for your specific item.
  5. AI negotiation — With one click, our ElevenLabs-powered voice assistant calls the local shops on your behalf. It describes the item, pitches the condition, and negotiates with the store owner using data-backed pricing scripts to secure the highest possible offer.
  6. Review offers — Responses from each store populate on your Offers page. You review the bids, pick the best one, and confirm.
  7. Ship & get paid — A custom shipping label is generated for your buyer, and you receive an SMS confirmation to track the status of your sale.

For the first time ever, capitalize on the full potential of your items with the least resistance possible.

Viability & Market Opportunity

The US secondary and resale market is valued at over $200 billion and growing year over year, fueled by sustainability trends and platforms like eBay, Mercari, and Facebook Marketplace. Yet the process of selling remains fragmented and manual, research the item, find a buyer, negotiate, ship. Each step is a friction point where people give up and leave money on the table.

FlipKit essentially collapses the entire workflow into a single interaction, eliminating the majority of friction points highlighted above. Our target users include casual declutterers cleaning out a home, garage sale flippers looking for underpriced inventory, estate sale liquidators processing items in bulk, and collectors who want to know the real value of what they own.

Revenue model: We take a small commission on completed sales brokered through the platform. A premium tier for power resellers unlocks batch scanning, priority AI negotiation calls, and auto-listing to multiple marketplaces. Affiliate partnerships with eBay, Mercari, and other platforms generate referral revenue when users list directly through FlipKit.

The path to revenue is built into the core product, every successful sale is a monetization event.

How we built it

Frontend: Next.js 16 with React 19, TypeScript, and Tailwind CSS. Authentication handled by Clerk. Deployed on Vercel.

Backend: FastAPI (Python) serving REST endpoints for image analysis, negotiation orchestration, and offer management. Deployed on Render.

Database & Storage: Supabase for persisting analyses, negotiation jobs, and store offers. Supabase Storage for secure image uploads with signed URLs.

Agent Pipeline (the brain): We built a multi-stage LangGraph pipeline that orchestrates the entire analysis:

  • Vision Node — Claude (Opus 4.6) analyzes the uploaded image and extracts structured product data: name, category, condition grade, notable features, and estimated age.
  • Pricing Swarm — Five parallel NVIDIA Nemotron (Nano 3-30b) workers each specialize in a different angle: online marketplace pricing, local marketplace pricing, condition impact analysis, market demand & trends, and pawn/resale shop valuations. They run simultaneously and return in seconds.
  • Shop Finder — Runs in parallel with the swarm, hitting SearchAPI.io to pull Google Maps data for nearby stores matched to the item's category.
  • Synthesis Node — Claude (Opus 4.6) combines all swarm analyses and shop data into a single actionable payload: price ranges, platform comparisons, a prioritized store list, and full negotiation scripts.

Voice Negotiation: ElevenLabs AI voice agent receives the synthesis payload and executes real phone calls to each store, using the generated negotiation scripts to secure the best price.

Design & Experience

FlipKit is built mobile-first because that's where the use case lives, you're standing at a garage sale or rummaging through your attic with your phone in hand. (youtube.com/watch?v=b-E2hOevym4&feature=youtu.be)

Challenges we ran into

  • Parallel agent orchestration — Coordinating five pricing workers plus a shop finder to run simultaneously and fan back into a single synthesis node required careful state management in LangGraph.
  • Keeping the pipeline under 60 seconds — Each individual agent call is fast, but chaining vision → swarm → synthesis end-to-end while maintaining quality meant aggressive parallelization and optimizing prompt sizes.
  • Accurate condition grading from a single photo — We tuned the vision prompts to be conservative and transparent about what's visible vs. what might be hidden.
  • Handling diverse item categories — We built dynamic search queries that adapt to the item category identified in the vision step.

Accomplishments that we're proud of

  • True multi-agent swarm — Five specialized pricing workers running in parallel, each analyzing the item from a different market perspective.
  • AI that makes real phone calls — Not just a chatbot. Our voice agent actually calls stores, describes items, and negotiates prices using data-backed scripts.
  • Full-stack product — Photo to identification to pricing to store matching to negotiation to shipping label. Every step of the secondary market hustle, automated.

What we learned

  • LangGraph for multi-agent orchestration — How to design graph topologies with parallel fan-out and fan-in patterns for real-time AI pipelines.
  • Multi-model architectures — Using Claude for high-reasoning tasks and NVIDIA Nemotron for fast parallel analysis creates a powerful combination of quality and speed.
  • Voice AI for real-world tasks — Building negotiation scripts that sound natural on a phone call is a very different challenge than writing chatbot responses.
  • Real-time pricing data is messy — Aggregating across platforms requires careful normalization.

What's next for FlipKit!

  • Multi-item batch scanning — Scan an entire box of items at once instead of one at a time.
  • Live auction mode — Let multiple stores bid against each other in real-time for your item.
  • Browser extension — Inline price checking while browsing Facebook Marketplace, eBay, or Craigslist.
  • Expanded marketplace integrations — Direct listing to platforms like eBay, Mercari, and Poshmark with AI-generated descriptions.
  • Buyer mode — Scan items at garage sales to instantly know if they're underpriced and worth flipping.

Built With

Share this project:

Updates