Fliplt

The secondhand market, but with AI doing all the work on both sides.

Fliplt is an autonomous resale agent that operates on both sides of the secondhand market. Point it at something in a thrift store and it identifies the item, pulls eBay sold comps, calculates your profit margin, and drafts a Depop listing — all before you leave the aisle. Tell it you want a rare sneaker and a swarm of agents hunts Depop, eBay, Mercari, and OfferUp, ranks the best listings, and sends negotiated offers to sellers automatically.

Table of Contents

• What It Does
• The Problem
• Our Solution
• Demo Flow
• Architecture
• Agent Roster
• Tech Stack
• Quick Start
• Local Setup
• Environment Variables
• Deployment
• Future Work

---

What It Does

• Sell mode — photograph an item at Goodwill, get a profit estimate and a ready-to-post Depop listing in under 60 seconds
• Buy mode — describe any niche item; agents search 4 platforms in parallel, rank results by value, and send optimized offer messages to haggle sellers down automatically
• Live agent dashboard — every item card shows real-time agent status as pipelines run, surfaced over SSE
• Autonomous negotiation — the negotiation agent reads market comps and crafts a per-seller offer message, then submits it via Browser Use without any manual input
• ASI:One integration — ask a natural-language question about your agents via Fetch.ai's ASI:One; it discovers and routes to your registered agents on Agentverse
• Configurable agent behavior — tune negotiation style (aggressive → passive), response delay, auto-reply, and reply tone per-account

---

The Problem

Resale is a grind on both sides.

Sellers walk into Goodwill, spot something that might be worth flipping, then manually check eBay sold listings, take photos on a bad background, type out a listing, set a price based on gut feel, and hope it sells. The research alone kills the in-store momentum.

Buyers hunting for something specific — a rare sneaker, vintage camera, limited-run streetwear — open Depop, eBay, Mercari, and OfferUp in separate tabs, scroll through stale listings, message sellers one by one, and almost always end up paying asking price because follow-up dies in the DMs.

The market is there. The friction is everything between the idea and the transaction.

---

Our Solution

Fliplt replaces both workflows with autonomous agents.

The sell loop:

1. Photograph the item → Vision agent identifies it and pulls a cleaned image
2. Research agent scrapes eBay sold comps via Browser Use
3. Pricing agent computes median price and profit margin
4. Listing agent populates a Depop draft form, ready to post in one tap

The buy loop:

1. Describe what you want → four search agents swarm Depop, eBay, Mercari, and OfferUp simultaneously
2. Ranking agent scores every listing by price-to-market and seller reliability
3. Negotiation agent generates a custom offer for each target seller and submits it
4. Conversations surface in the item detail view as replies come in

Every agent runs as a registered Fetch.ai uAgent on Agentverse, implementing the Chat Protocol so ASI:One can discover and orchestrate them. The mobile app talks to the same pipeline over FastAPI + SSE for low-latency demo reliability.

---

Demo Flow

1. Dashboard — your resale portfolio at a glance

P&L chart, net profit, selling items with margin estimates, and buying items with offer statuses. One screen shows you whether the agents are working.

2. Configure your agents

Per-account settings for negotiation style, response delay, auto-reply, and reply tone. Set it aggressive when you're hunting, dial it back when you want to appear casual.

3. Watch the negotiation happen

The item detail view shows the market price range, active platforms, and the live conversation thread as sellers respond to agent-sent offers.

4. Talk to your agents via ASI:One

Ask natural-language questions through Fetch.ai's ASI:One interface. It routes your query to the right registered agent on Agentverse and returns the response.

---

Architecture

flowchart TD
    App[Expo Mobile App] -->|JWT + HTTP| API[FastAPI Backend]
    App -->|SSE| Stream[Event Stream]
    App -->|Chat Protocol| ASI[ASI:One / Agentverse]

    API --> SellPipeline
    API --> BuyPipeline

    subgraph SellPipeline [Sell Pipeline]
        V[VisionAgent] --> E[EbayResearchAgent]
        E --> P[PricingAgent]
        P --> D[DepopListingAgent]
    end

    subgraph BuyPipeline [Buy Pipeline]
        DS[DepopSearchAgent] --> R[RankingAgent]
        ES[EbaySearchAgent] --> R
        MS[MercariSearchAgent] --> R
        OS[OfferUpSearchAgent] --> R
        R --> N[NegotiationAgent]
    end

    API -->|verify ownership + read runs| DB[(Supabase)]
    API --> SM[SessionManager]
    SM -->|persist events + results| DB

    ASI -->|Mailbox| Agents[10 uAgents on Agentverse]
    Agents -->|HTTP - delegates to pipeline| API

Loading

Key design decisions

• Two front doors, one agent system — FastAPI runs the same 10-agent pipeline in-process for mobile; Fetch.ai/Agentverse exposes the same agents for ASI:One discovery and natural-language orchestration
• SSE for real-time progress — SessionManager holds state in memory with async pub/sub queues; each agent step publishes events the client streams live
• 3-tier search resolution — each buy search agent tries httpx first, falls back to Browser Use, then to deterministic fallback data. Resilient under demo conditions
• Sell pipeline pause points — pipeline suspends after vision (low-confidence correction flow) and after listing draft (user reviews before posting). Both resume cleanly
• Supabase RLS — all data scoped to auth.uid(); backend verifies item ownership before any pipeline operation

---

Agent Roster

All 10 agents registered on Agentverse, implementing Chat Protocol, discoverable by ASI:One.

#	Agent	Pipeline	Responsibility	Key APIs
1	VisionAgent	SELL	Item identification + clean photo	Gemini Vision, Nano Banana
2	EbayResearchAgent	SELL	eBay sold comp scraping	Browser Use
3	PricingAgent	SELL	Median price + profit margin	Gemini
4	DepopListingAgent	SELL	Depop draft form population	Browser Use
5	DepopSearchAgent	BUY	Active Depop listing search	Browser Use
6	EbaySearchAgent	BUY	Active eBay listing search	Browser Use
7	MercariSearchAgent	BUY	Active Mercari listing search	Browser Use
8	OfferUpSearchAgent	BUY	Active OfferUp listing search	Browser Use
9	RankingAgent	BUY	Score + rank all listings	Gemini
10	NegotiationAgent	BUY	Generate + submit offer per seller	Browser Use, Gemini

---

Tech Stack

Mobile

• React Native + Expo SDK 54
• expo-router v6 (file-based routing)
• Supabase JS client (auth + database)
• NativeWind (Tailwind for RN)

Backend

• FastAPI + Uvicorn
• Pydantic v2 (strict agent I/O schemas)
• Browser Use + Patchright (Chromium automation)
• Google Gemini (Vision, text generation)

Infrastructure

• Supabase (Postgres + Auth + Storage + RLS)
• Fetch.ai uAgents + Agentverse Mailbox
• Render (backend deployment)

AI

• Gemini Vision — item identification from photo
• Gemini text — pricing, ranking, negotiation message generation
• Nano Banana — professional product photo cleanup
• Browser Use — live web scraping and form submission across all 4 platforms
• ASI:One — natural-language orchestration across the agent network

---

Quick Start

# Backend
make install
make run

# Frontend (separate terminal)
cd frontend && npm start

Requires .env with at minimum GOOGLE_API_KEY and SUPABASE_URL / SUPABASE_SERVICE_ROLE_KEY. See Environment Variables.

---

Local Setup

Backend

make install          # creates .venv and installs requirements.txt
make check            # runs tests + compile check
make run              # starts uvicorn on port 8000

For live Browser Use flows (scraping + form submission):

# Chromium is installed by make install via patchright
# Warm browser profiles before demo:
./.venv/bin/python scripts/create_browser_profile.py
./.venv/bin/python -m backend.browser_use_runtime_audit   # verify setup

To skip Browser Use and use deterministic fallback data:

BROWSER_USE_FORCE_FALLBACK=true make run

Frontend

cd frontend
npm start             # Expo dev server (choose platform from menu)
npm run ios           # iOS simulator
npm run android       # Android emulator
npm run web           # web browser

Backend URL defaults to http://localhost:8000. Override with EXPO_PUBLIC_API_URL in frontend/.env.

Fetch.ai agents (optional)

Requires Python 3.12 or 3.13 (uagents 0.24.0 is incompatible with 3.14):

make venv-fetch
set -a && source .env && set +a
make run-fetch-agents         # all public Agentverse-facing agents
make run-fetch-agent AGENT=vision_agent   # single agent with inspector URL

---

Environment Variables

Variable	Default	Purpose
GOOGLE_API_KEY	(required)	Gemini Vision + Browser Use
NANO_BANANA_API_KEY	(required)	Gemini key for AI photo analysis
SUPABASE_URL	(required)	Supabase project URL
SUPABASE_SERVICE_ROLE_KEY	(required)	Supabase service role key
SUPABASE_JWT_SECRET	(required)	JWT verification secret
EBAY_APP_ID, EBAY_CERT_ID	(none)	eBay Browse API credentials
INTERNAL_API_TOKEN	dev-internal-token	Auth for internal event endpoint
AGENT_EXECUTION_MODE	local_functions	local_functions or http
BROWSER_USE_FORCE_FALLBACK	false	Skip Chromium, return mock data
BROWSER_USE_GEMINI_MODEL	gemini-2.0-flash	Model for Browser Use flows
FETCH_ENABLED	false	Enable Fetch.ai uAgents layer
SCHEDULER_INTERVAL	300	Seconds between background scheduler polls
APP_BASE_URL	http://localhost:8000	Stream/result URL construction
ALLOWED_ORIGINS	*	CORS origins (comma-separated)

---

Deployment

Deployed on Render (render.yaml). Build installs Chromium for live Browser Use:

pip install -r requirements.txt && python -m patchright install chromium

Notes:

• Headed Chromium requires a paid Render instance — free tier is not reliable for live Browser Use demos
• Set GOOGLE_API_KEY, SUPABASE_SERVICE_ROLE_KEY, and INTERNAL_API_TOKEN as Render secrets, not in render.yaml
• BROWSER_USE_FORCE_FALLBACK=true gives a fully functional demo without Chromium if needed
• Run the validation harness before demos: make verify-browser

---

Future Work

• Parallel buy search — fan out all 4 platform searches simultaneously (groundwork exists in the Fetch.ai path)
• In-store scan mode — barcode + camera scan at Goodwill with instant margin estimate
• Cross-platform listing — post to Depop, eBay, and Mercari in one tap
• Offer tracking — persistent negotiation threads with seller reply detection and counter-offer handling
• Web app — the backend is platform-agnostic; a Next.js frontend is a natural next step

---

Built at DiamondHacks 2026 · UCSD · April 5–6