Evidence scattered across systems
Order, payment, fulfillment, refund, and policy detail live in different places, and disputes have a hard reply deadline.
Working build
Proofly: AI Chargeback Dispute Copilot
A self-hosted dispute copilot that reads a merchant's order database and policy documents, then recommends contesting or accepting a chargeback with plain-English reasoning, policy citations, and a downloadable evidence report.
StatusWorking build
PlatformWindows
AI LayerDispute review
FocusChargebacks
A working build by CyphronTech that is not yet live. No usage or client results are claimed.
Overview
Turning scattered order evidence into a reviewable recommendation.
Proofly was built around a narrow, high-stakes task: a merchant has around 48 hours to answer an "item not received" chargeback, and the proof is spread across the order database, a policy PDF, and staff memory. The app reads an order by ID, pulls the matching policy clause, and asks an AI model to weigh the two.
A deterministic safety net cross-checks the AI's recommendation against the merchant's own records before anything is shown, and routes the case to manual review instead of trusting a confident but wrong answer.
What the build focused on
Read-only AI review with a safety net
A read-only, allowlisted connector reads the order; the AI reviews it against the cited policy clause; a deterministic check overrides the AI if it contradicts the merchant's own records.
Fewer disputes lost to guesswork
Merchants get a plain-English recommendation and a downloadable evidence report instead of digging through tables by hand.
Scope
A dispute workflow built to be checked, not just trusted.
Case review
Order ID in, plain-English summary out: what was ordered, delivery status, refund history, and the AI's contest/accept recommendation with a confidence score.
Policy citations
The AI quotes the merchant's actual refund or shipping policy, and every quoted clause is verified against the document text before it is shown.
Manual review queue
Cases the safety net disagrees with, or the AI is unsure about, are flagged for a human to decide instead of auto-approved.
Evidence report
An approved case exports as a PDF, with a fingerprint hash proving it was not edited after approval, ready to send to the payment processor.
AI workflow
AI that is checked against the merchant's own records.
The AI layer reads order, payment, fulfillment, and refund evidence alongside the cited policy clause and recommends contest or accept. A deterministic rule layer then re-checks that recommendation against facts the connector itself read, and overrides the AI when the two disagree.
- Evidence reviewWeighs delivery status, refund history, and return requests against the matching policy clause.
- Cited, verified quotesEvery policy quote is checked against the source document before it reaches the merchant.
- Safety-net overrideA refund already issued or an open return request routes the case to manual review, no matter what the AI recommends.
- Bring-your-own modelWorks with Anthropic, OpenAI-compatible APIs, or a local model, with the provider key encrypted at rest.
Implementation direction
Built to run on the merchant's own machine.
The product direction covers a read-only, allowlisted database connector with schema-validated queries, a setup wizard that maps a merchant's own tables and columns, keyword-based policy retrieval with no vector database required, a full audit trail, and packaging as a single Windows installer with an embedded database.
Spring BootReactTypeScriptPostgreSQLAnthropicWindows
Need AI that's checked against your own data?
We can design read-only, audited AI review workflows for the high-stakes decisions your team makes under a deadline.