SupportDeskHelper — ResolveLoop

AI customer support for ecommerce that automates safe, policy-based replies and verifies whether promised resolutions were actually completed.

Comment

Inspiration

After running my own Shopify stores, I noticed a gap in customer support: teams could respond quickly and close tickets, but a closed ticket did not prove that the customer’s problem had actually been solved.

Most support tools focus on generating faster replies and reducing resolution time. I wanted to build a system that could also examine what happened after the reply—especially when a business promised a refund, replacement, reshipment, tracking update, or follow-up.

That idea became SupportDeskHelper. During OpenAI Build Week, I built ResolveLoop, a verification layer designed to evaluate whether customer commitments were actually fulfilled.

What it does

SupportDeskHelper — ResolveLoop is a customer support platform for ecommerce businesses, developed with GPT-5.6 and Codex.

It brings customer conversations, business policies, customer history, Shopify order context, internal notes, follow-ups, team workflows, and ticket management into one workspace.

SupportDeskHelper uses each business’s own policies and verified ticket context to create relevant reply drafts instead of relying on unsupported assumptions.

For safe, policy-backed requests, businesses can enable Auto Reply to send an immediate response. Sensitive requests remain under human control, while agents can review drafts, adjust the tone, attach files, add private notes, or respond manually.

Shopify order information is displayed inside the ticket when available. If an order exists but the customer’s ownership cannot be confirmed, SupportDeskHelper requires manual verification before allowing private order details to be used in a response.

ResolveLoop

ResolveLoop is the main feature I built during OpenAI Build Week.

Most support platforms stop when an agent marks a ticket as resolved. ResolveLoop continues after that point.

It examines the completed support conversation to determine:

  1. What the customer requested
  2. Whether the support team made a concrete commitment
  3. What evidence would be required to prove completion
  4. Whether the available ticket or Shopify evidence supports that commitment

A commitment might include issuing a refund, sending a replacement, reshipping an order, providing tracking information, or following up by a specific date.

ResolveLoop then produces a structured result inside the ticket:

  • Verified — the stored evidence supports that the commitment was completed
  • Pending — a commitment was made, but completion is still awaiting evidence
  • At risk — the deadline is approaching and the required evidence is missing
  • Unverified — the deadline has passed without sufficient evidence
  • Ambiguous — the conversation or commitment is unclear
  • Not applicable — no supported agent commitment was detected

ResolveLoop also includes a Why this result? explanation showing:

  • The promise that was reviewed
  • The source message
  • The evidence considered
  • The reason for the decision
  • Any evidence limitations
  • The recommended next action

When the evidence is incomplete, ResolveLoop does not guess. It explains what is missing and communicates uncertainty clearly.

Closing a ticket is not the same as solving the customer’s problem.

How we built it

SupportDeskHelper was developed with GPT-5.6 and Codex supporting the product and full-stack engineering process.

Codex helped audit, implement, test, and improve areas including:

  • Authentication and workspace access
  • Owner and agent permissions
  • Ticket workflows and customer records
  • Internal notes and conversation history
  • Policy-grounded reply logic
  • Auto Reply safety guardrails
  • Shopify synchronization and customer verification
  • Outbound and inbound email handling
  • Notifications and per-user unread state
  • Analytics, debugging, and production fixes

During OpenAI Build Week, I focused specifically on designing and implementing ResolveLoop.

ResolveLoop uses deterministic TypeScript logic to examine completed conversations, identify whether a concrete support-agent promise was made, classify the commitment, evaluate stored ticket and Shopify evidence, and explain the resulting verification status.

GPT-5.6 and Codex were used extensively for:

  • Product design and requirements
  • Acceptance criteria and edge cases
  • Architecture decisions
  • Implementation planning
  • Code generation and refactoring
  • Debugging and repository audits
  • Test creation and validation
  • Safety and privacy reviews
  • Documentation and submission preparation

The current production verification engine does not call the OpenAI API at runtime. GPT-5.6 and Codex were used as engineering, reasoning, testing, and auditing partners throughout Build Week.

The application converts each verification result into structured data displayed directly inside the ticket, allowing agents to understand the detected promise, supporting evidence, limitations, decision reason, and next action.

The application is built with:

  • Next.js
  • React
  • TypeScript
  • Supabase
  • PostgreSQL
  • Shopify
  • Resend
  • Stripe
  • Tailwind CSS
  • Vercel
  • GitHub

Challenges we ran into

The hardest challenge was separating what the system could infer from what the available evidence could actually verify.

A customer request is not automatically a business promise. If a customer asks for a refund, that does not mean the business agreed to issue one.

A resolved ticket status is also not proof that a promised action was completed.

ResolveLoop therefore had to distinguish between:

  • The customer’s request
  • The support agent’s actual commitment
  • The deadline associated with that commitment
  • The evidence required to prove fulfillment

Another challenge was handling incomplete or contradictory conversations. The system needed to recognize uncertainty instead of producing confident conclusions that the evidence did not support.

Shopify data also introduced privacy and identity-verification challenges. An order number may exist even when the customer’s ownership cannot be confirmed. In those cases, SupportDeskHelper requires manual verification instead of exposing private order information or using it in a customer-facing reply.

The final challenge was keeping ResolveLoop isolated and safe. It can analyze stored information and explain a result, but it cannot issue refunds, create replacements, modify orders, send customer emails, or change ticket status automatically.

Accomplishments that I’m proud of

The main accomplishment is building a working ecommerce support platform that goes beyond fast replies and ticket closure.

SupportDeskHelper combines:

  • Ticket and conversation management
  • Business policies
  • Customer history
  • Shopify order context
  • Email delivery
  • Team permissions
  • Internal notes
  • Follow-up tracking
  • Analytics
  • Controlled reply automation

ResolveLoop adds a new accountability layer by evaluating whether customer commitments were actually completed.

I am particularly proud that ResolveLoop:

  • Separates customer requests from agent commitments
  • Requires evidence before marking a promise as verified
  • Produces structured and explainable results
  • Recognizes missing or contradictory information
  • Avoids presenting assumptions as facts
  • Keeps sensitive actions under human control
  • Fits directly into the existing ticket workflow
  • Provides a clear next action when verification fails

The result is not only a faster support platform. It is a system designed to make customer outcomes measurable, explainable, and accountable.

What we learned

I learned that trustworthy automation is not only about producing a useful response. It is also about knowing when the available information is not strong enough to support a conclusion.

The most reliable workflows combine structured data, deterministic rules, clear evidence, explicit limitations, and human oversight.

I also learned that uncertainty should be treated as useful information rather than hidden. A system that clearly says, “This cannot yet be verified,” can be more valuable than one that produces a confident but unsupported answer.

GPT-5.6 and Codex were especially valuable as engineering partners. They helped turn a broad product idea into specific requirements, edge cases, tests, safety boundaries, implementation decisions, and production-ready documentation.

ResolveLoop changed the direction of SupportDeskHelper. The project began as a way to help ecommerce teams respond faster. It is now becoming a system that also helps businesses evaluate whether customers actually received the outcomes they were promised.

What’s next for SupportDeskHelper — ResolveLoop

The next step is to expand ResolveLoop with stronger evidence sources, automated follow-up tracking, and clearer resolution histories.

Planned improvements include:

  • Automated follow-up checks
  • Additional Shopify refund and fulfillment evidence
  • Live evidence refresh during verification
  • Inbound email ticket creation
  • Stronger customer identity verification
  • Automated customer follow-ups
  • Resolution analytics
  • Commitment and deadline tracking
  • Reporting on verified, pending, at-risk, and unverified outcomes

The long-term goal is to help ecommerce businesses move beyond measuring response time and ticket-closure rates—and start measuring whether customer problems were actually solved.

Built With

Share this project:

Updates