ClaimIt

Claimit Dashboard

You bought something. The price dropped. You deserve a refund.

ClaimIt is a multi-agent AI system that automates post-purchase price-match refunds. It monitors prices across retail, airline, and hotel platforms, detects refund eligibility under each platform's own policy, and generates the exact claim material the platform requires — email, chat script, in-store guide, or self-service walkthrough. For email claims, it can send directly from your Gmail after approval.

Current submission: Google Cloud Rapid Agent Hackathon — Agent Builder + Gemini track, MongoDB MCP partner integration.

Gmail connection is invite-only. ClaimIt's Google OAuth app is still in Google's "Testing" status (sensitive Gmail scopes require Google verification before public release), so only allowlisted test users can link their mailbox up to Google's 100-test-user cap. To try the Gmail receipt-detection and claim-sending features, contact the ClaimIt team to be added as a test user first.(Please contact claimitbeta@gmail.com for permission)

Why ClaimIt Exists

Nearly every major retailer, airline, and hotel will refund you when the price drops after you buy, but almost no one ever claims it. Industry estimates put unclaimed price-protection refunds in the U.S. at $10B+ a year. The refunds are legitimate; the process is just too fragmented to run by hand — claim windows run from ~14 days (retail) to as little as 24 hours (hotels), and every platform demands a different submission format.

ClaimIt automates the watching, the policy interpretation, and the writing — and handles submission where the platform allows it.

How It Works

Ingest — Gemini parses purchase receipts from Gmail or uploads into structured data
Monitor — Prices tracked on a cadence-aware schedule (6h → 1h → 15min as windows close), with screenshot evidence at every check. Product URLs are auto-resolved from receipt data when missing
Draft — Claim materials generated in the format each platform actually accepts, self-evaluated by Gemini before surfacing to the user
Approve — Three-pane review UI: draft + evidence + AI assistant. Edit if needed, then approve
Submit — Email claims sent from your Gmail after approval. Other types provide copy-ready scripts, printable guides, or step-by-step walkthroughs

Output Type	On Approve	Example Platforms
Email	Sent from your Gmail after approval	Best Buy, Delta, Hilton
Chat Script	Copy messages into live chat	Amazon, Target
In-Store Guide	Print and bring to the store	Target, Home Depot
Self-Service Walkthrough	Follow step-by-step instructions	Southwest, United

→ Claim Output Types

Tech Stack & MCP Integrations

Core Technologies

Layer	Technologies
AI	Google Cloud Agent Builder · Gemini · Vertex AI
Data	MongoDB Atlas · Elasticsearch
Events	Google Cloud Pub/Sub · Cloud Scheduler
Compute	Google Cloud Run (8 services)
Frontend	Next.js · TypeScript · Tailwind CSS · shadcn/ui
Backend	Python · FastAPI · Pydantic v2
Auth	Firebase Authentication · Gmail OAuth 2.0
Observability	Arize Phoenix · OpenTelemetry
Infra	Terraform · GitHub Actions

→ Tech Stack Details

Three MCP Integrations

Agents access external systems through the Model Context Protocol. Each MCP is fronted by Gemini-safe FunctionTool wrappers — raw MCP schemas use const/oneOf constructs that Gemini rejects, so each tool is wrapped with a simple, parseable signature. Auth is injected via httpx request event hooks (Agent Engine silently drops client-level auth).

MCP	Purpose	Transport	Agents
MongoDB	Exact-match policy lookup by platform name	Streamable HTTP → Cloud Run (OIDC)	All four agents
Phoenix	Read OTel traces and spans — explain claim reasoning, surface self-eval scores	stdio→HTTP bridge → Cloud Run (OIDC)	All four (claim reasoning: Assistant + Claim)
Elastic	Full-text policy search across all platforms	Streamable HTTP → Kibana (ApiKey)	Assistant only

MongoDB MCP handles precise, named-platform lookups (search_platform_policy("best_buy")). Elastic MCP complements it with broad, natural-language search (search_policies_fulltext("return window electronics")). Phoenix MCP lets the Assistant explain why a claim was drafted the way it was by reading the Gemini call traces.

Tenancy is enforced in the wrapper layer — collection/filter constraints are hard-coded, not model-controlled. Secrets are read from environment at request time, never baked into cloudpickle or Terraform state.

→ MCP Integrations

Architecture

Four AI agents collaborate through a Google Cloud Pub/Sub event bus. Each agent has a single responsibility. No agent calls another directly — they communicate exclusively through events.

ClaimIt agent architecture

Agent	Responsibility
Ingest	Receipt parsing via Gemini; Gmail watch integration; low-confidence extractions trigger human confirmation
Monitor	Cadence-aware price surveillance with screenshot evidence capture; auto-resolves product URLs from receipt data via ScraperAPI when missing
Claim	Four output types; Gemini self-critique on clarity, tone, accuracy, completeness; Gmail Send API for email claims
Assistant	Mode A: general help · Mode B: claim-scoped with full context, tool access, and reasoning trace visibility. Invoked through the API Gateway and grounded in claim context plus trace metadata

→ Architecture

MongoDB Atlas as Shared Context Layer

MongoDB Atlas stores the full claim lifecycle across six collections, accessed by agents through MCP tool calls:

Collection	Purpose
`users`	Accounts, Gmail OAuth state, send-mode preferences
`purchases`	Extracted purchase records with confidence scores and resolved product URLs
`price_history`	Timestamped price observations with evidence references
`policies`	Platform refund policies — windows, exclusions, submission channels
`claims`	Draft versions, approval history, submission state, outcomes
`conversations`	Assistant chat threads and agent session references

→ Data Model Reference

Technical Highlights

Event-driven multi-agent orchestration on Google Cloud Agent Builder with Pub/Sub — agents scale and fail independently
Three MCP integrations (MongoDB, Phoenix, Elastic) with Gemini-safe wrappers, event-hook auth, and wrapper-level tenancy enforcement
Four distinct claim output types matched to real platform submission processes, not generic templates
Gemini self-evaluation loop grades drafts on a rubric before surfacing them to users
Recursive observability via Arize Phoenix — the Assistant queries Phoenix to explain claim decisions, and that query is itself traced
Gmail integration for both inbound receipt detection and outbound claim sending from the user's own mailbox

Platform Coverage

26 platforms across three categories (24 with active price-protection policies):

Category	Platforms
Retail (15)	Best Buy · Target · Dick's · JCPenney · Staples · Macy's · Nordstrom · Newegg · Costco · Dell · Home Depot · Lowe's · Crutchfield · ~~Amazon~~ · ~~Walmart~~
Airlines (6)	Southwest · Delta · United · JetBlue · Alaska · American
Hotels (5)	Hilton · Marriott · Hyatt · IHG · Wyndham

Two platform (Best Buy and Target) run fully live end-to-end. Amazon and Walmart are tracked but currently inactive (no post-purchase price-match program). The remaining use source-linked policy fixtures and seeded data to demonstrate scalable coverage. Adding a new platform requires only a policy document and a price-check adapter.

→ Platform Coverage

Security & Privacy

Gmail OAuth is configured to request only the scopes needed for receipt detection and claim sending — see Security Model for the full scope list
OAuth tokens stored in Google Cloud Secret Manager, not in application databases
No email is ever sent without explicit user approval (or a user-configured auto-send with 5-minute cancellation window)
MCP tool wrappers enforce tenancy at the code level — the model cannot pivot queries to other users' data
Current deployment is single-tenant; multi-tenant isolation is planned

→ Security Model

Limitations

2 of 26 platforms run live end-to-end; the remaining 24 use source-linked policy fixtures with simulated price monitoring
Outcome tracking is manual — users report whether the merchant approved or denied the claim
Per-user quotas for Gemini calls and scraper requests are designed but not yet enforced
Gmail connection is invite-only. ClaimIt's Google OAuth app is still in Google's "Testing" status (sensitive Gmail scopes require Google verification before public release), so only allowlisted test users can link their mailbox up to Google's 100-test-user cap. To try the Gmail receipt-detection and claim-sending features, contact the ClaimIt team to be added as a test user first.

Future Plans

Full live coverage — extend real end-to-end monitoring from Best Buy to all 24 platforms

Open Gmail to everyone — complete Google OAuth verification to remove the invite-only test-user limit
Close the loop — auto-redraft on low self-eval scores and automate outcome tracking (today it's manual)
Scale out — multi-tenant isolation, then a browser extension and mobile app for receipt capture

Team

Name	Role	Focus
Erdun E	PM · Frontend	Product direction, Next.js UI, demo production
Raj Kavathekar	Agent AI Engineer	Ingest Agent, Monitor Agent, Assistant Agent, data pipelines
Kewen Chen	Agent AI Engineer	Claim Agent, prompt engineering
Qingyuan Wan	Backend · Infrastructure	API Gateway, shared packages, Terraform, CI/CD, MCP integrations