ResolveIQ — Enterprise AI Case Resolution

Overview

ResolveIQ is an enterprise AI platform for case and incident resolution. It takes a messy, high-stakes support case and turns it into a resolution recommendation that cites the policies and prior incidents behind it, scores the risk, and stops at a human approval gate before anything is actioned. The tagline on the app says it plainly: evidence in, decision yours.

It is a real deployment running on AWS with synthetic data. You can open the live demo, pick a role, and walk a case end to end without an account.

The metrics below come from a synthetic golden set built to exercise the system, not from a customer production environment. I frame them as demonstrated behavior, not delivered business ROI.

Business problem

Support and operations teams face the same three problems at once:

• High-risk cases hide inside ordinary ones. The case that breaches an SLA or churns an account looks like every other ticket until it is too late.
• The answer usually exists already, in a policy, a runbook, or a senior engineer's memory. Finding it and trusting it is the job.
• An AI assistant can help with both, but only if a reviewer can see why it said what it said and stays in control of what happens next.

Target users

• Case specialists who need a defensible next step, fast.
• Operations leaders who need visibility into risk and consistency.
• AI and architecture reviewers who need to audit how a recommendation was formed.

The live demo lets you assume each of these roles so you can see the same case from different vantage points.

Architecture

ResolveIQ is serverless and AWS-native. A case moves through an orchestrated pipeline with a human decision at the end.

1. A case enters and is grouped and enriched, then risk-scored by an ML inference step.
2. A retrieval step pulls relevant policies and similar prior incidents from an Amazon Bedrock Knowledge Base backed by an S3 vector store.
3. Retrieved evidence, the risk score, and case context are composed into a grounded recommendation with inline source citations.
4. The recommendation is presented for a human to approve, modify, or escalate. Nothing acts on a customer automatically.
5. Every step, including the human decision, is written to an audit trail.

Architecture walkthrough

• Amplify hosts the Next.js frontend; Cognito guards an API Gateway HTTP API.
• Domain-grouped Lambda functions handle the application logic, orchestrated by AWS Step Functions for the multi-step resolution flow.
• Amazon Bedrock Knowledge Bases provide retrieval over an S3 vector store, with Bedrock Guardrails on the generation step.
• A machine learning step produces the risk score that routes attention to the cases that need it.
• DynamoDB holds case state and the audit history; EventBridge and CloudWatch cover events and observability.
• The whole stack is defined in AWS CDK, so the environment is reproducible and every change is reviewable.

Important design decisions

• Retrieval before generation, always. The model composes from retrieved evidence and cites it. If the evidence is weak, the recommendation says so rather than guessing.
• Human-in-the-loop by default. The system drafts and recommends; a person decides. That line is the difference between a copilot and an agent acting on a customer.
• Role-based access. Case specialist, operations leader, and reviewer see different controls, which keeps approval authority where it belongs.
• Scale to zero with per-run cost visibility. The design shows what a single resolution costs, because cost is a first-order concern for anyone putting this in production.

Security and governance

• Guardrails on generation, source attribution on every claim, and an approval gate before any downstream action.
• An immutable audit trail records inputs, retrieved sources, the risk score, and the human decision.
• Least-privilege IAM and Cognito-guarded APIs; no long-lived credentials in the app.

Evaluation & success metrics

• Retrieval is measured against a labeled set to confirm the right evidence surfaces for representative cases.
• Grounding checks flag any recommendation not supported by its citations.
• The risk model is monitored with standard classification metrics.

Challenges

• The hard part was not retrieval. It was making the generation step refuse to answer confidently when the evidence did not support it.
• Designing the approval gate so it felt like leverage for a reviewer, not a speed bump, took more iteration than the model work did.

Lessons learned

• The gap between a good demo and a deployable system is mostly governance, evaluation, and human-in-the-loop design, not model quality.
• Citations are what make an AI recommendation reviewable, and reviewability is what makes it shippable.

Links

• Live demo: resolveiq.mikemiller.ai