Demo environment — do not upload sensitive or real company data. Data is reset periodically.
Interview demo only · built for LaunchGood's Applied AI Engineer application · not affiliated with LaunchGood

AI assisted campaign review,
with humans in control.

Review teams face more campaigns than reviewers can read with care. This demo shows an AI assisted workflow that prepares evidence linked packets so humans decide with confidence.

LaunchGood's public core business is helping fundraisers create campaigns and connect with donors. Campaign Operations Copilot demonstrates a representative internal workflow adjacent to that core: preparing submitted campaigns for consistent human review backed by evidence. This is a product hypothesis based on LaunchGood's public submit and review model. It does not reflect knowledge of internal workflows or confirmed pain points.

Start the guided demoSee how it works
30 deterministic fixtures1.00 contradiction detection1.00 missing evidence recall0.875 citation correctnessfull results →
Contradiction detected
Narrative claims $2,800 needed · School invoice totals $2,000
Source → Green Valley Academy Term 1 invoice see it in the demo →
Guided demo

Five minutes, five steps.

No signup required. Click a role below to start a seeded session and experience the full campaign review workflow.

1

Open the campaign

See prioritized campaign cases and understand why each needs review.

2

Inspect evidence

Read campaign narratives alongside supporting documents. The AI extracts claims and links every finding to its source.

3

Review findings

Examine cited facts, policy references, flagged contradictions, missing evidence, and confidence scores.

4

Authorize the decision

Edit the draft, preview an organizer message, then confirm: request info, escalate, approve, or reject. Only you can authorize.

5

Verify the record

Every retrieval, tool call, AI output, edit, and action is recorded in an auditable trace with latency and cost.

Try the demo

Pick a role to sign in

Each button signs you in as a real seeded user with a review queue, campaign cases, and an audit trail. All data is synthetic and fictional.

Reviewer
Campaign Review · Reviewer role

Open the review queue, run analysis on campaign evidence with AI assistance, inspect findings linked to sources, edit draft follow up questions, and request information from organizers.

Senior Reviewer
Campaign Review · Senior Reviewer role

All reviewer capabilities plus approve or reject reviewed cases, handle escalated campaigns, and supervise team decisions.

Policy Specialist
Campaign Review · Policy Specialist role

Review sensitive policy cases escalated from the review team. Apply policy specific checks before final decisions.

Administrator
Campaign Review · Administrator role

Manage demo data, view evaluation results, inspect audit traces and security events for all campaign reviews.

All data is synthetic and fictional.

Human/AI boundary

AI prepares the review. People make the decisions.

✓AI can

  • •Extract claims and facts from narratives and evidence
  • •Connect every finding to its source document
  • •Compare dates, amounts, identities, and intended fund use
  • •Retrieve and cite relevant policy passages
  • •Identify missing evidence and contradictions
  • •Summarize ambiguity and propose follow up questions
  • •Estimate confidence and abstain when evidence is insufficient

✗AI cannot, only humans

  • •Decide whether evidence is adequate or a campaign is truthful
  • •Make or approve a policy specialist judgment
  • •Approve, reject, escalate, or request information
  • •Send a message or change campaign state
  • •Authorize any action — model output is never authorization

Model output is never authorization. Every permission and state transition is checked again on the server.

How every action flows
AI proposesPolicy checksHuman approvesTool executesResult verifiedAudit logged
Known limits

Designed to fail in the open.

  • Abstains when evidence is insufficient instead of guessing
  • Contains injected instructions inside untrusted evidence
  • Escalates sensitive and cross border cases to specialists
Architecture

Built on Secure Enterprise Agent Gateway

  • Tenant isolated Postgres with RLS keyed off tenant_memberships.
  • Access filtered pgvector retrieval. Restricted tags never enter the prompt.
  • Tool gateway with role gates, schema validation, and policy decisions on every call.
  • Campaign state machine enforced on the server, shared by UI and agent.
  • Prompt injection defense in depth: untrusted block, pattern detector, audit log.
  • Full audit traces for retrievals, tool calls, blocked events, latencies, and cost.
Evaluation

Measured on 30 deterministic fixtures.

Contradiction detection 1.00 · missing evidence recall 1.00 · citation correctness 0.875 · claim extraction recall 1.00. full results →

Security

Isolation and policy checks on every call.

  • Tenant isolated data access enforced by row level security
  • Every tool call checked against the caller role policy
  • Untrusted content never becomes instruction
Observability

Every step traced.

Retrievals, tool calls, blocked events, latency figures and cost figures recorded per action and visible in the audit trail.

Extensible platform

One gateway, many workflows

The same tool registry, policy engine, RAG layer, authorization, and audit system can support additional internal modules.

HR policy Q&A

Answer policy questions using retrieval scoped by access rules.

PTO and leave workflows

Submit, approve, and track time off requests.

Support ticket triage

Route, prioritize, and draft responses.

Finance reconciliation

Match transactions and flag exceptions.

Internal knowledge search

Search over internal documents using natural language.

Campaign review is the focused application. The gateway architecture supports additional workflows by registering new policies, tools, roles, and approvals.