AuditFlow: Traceable Multi-Agent Audit System
Enterprise finance teams waste hours reconciling transaction data scattered across CRM, ERP, and Finance systems. The same deal appears differently in each — contract amounts don't match invoices, payments don't reconcile with bookings. Today this means manual spreadsheet work with no audit trail. AuditFlow solves this with 6 specialised agents across 4 layers, all coordinated through Band's room and @mention system. The Router Agent receives natural language queries, classifies intent, and dispatches to system agents via @mention. Three parallel System Agents (CRM, ERP, Finance) each query their own database and return structured data plus the business rules that govern it — payment terms, FX policies, invoicing schedules. The Reconciliation Agent aligns outputs and identifies every field-level discrepancy. The Root-Cause Agent diagnoses each discrepancy using those business rules, producing a confidence-scored verdict: normal (explained by policy), anomaly (unexplained, needs human review), or watch (conditional, needs follow-up). Every handoff is a real Band @mention. The full room conversation is the audit trail — every reasoning step timestamped and inspectable. Band isn't a wrapper here; it's the coordination layer that makes agent synchronisation and shared context real. The core insight: a single LLM can't reliably reconcile data across systems with conflicting business rules. When it fails, there's no way to locate which reasoning step went wrong. AuditFlow's role-based design makes every failure localised and traceable. Validated across 10 scenarios including instalment payments, FX conversions, partial refunds, entity name variations, and missing fields. AuditFlow reduces a multi-hour manual process to a structured, auditable workflow — with a decision trace that satisfies compliance requirements no black-box output ever could.
Category tags:
"Application of technology: Six-agent 4-layer architecture: Router (NLP intent classification + dispatch) → parallel CRM/ERP/Finance System Agents (each with own database + business rules) → Reconciliation Agent (field-level discrepancy alignment) → Root-Cause Agent (confidence-scored verdicts: normal/anomaly/watch). Validated across 10 scenarios (instalment payments, FX, partial refunds, entity name variations, missing fields). Presentation: Outstanding — "the full room conversation is the audit trail" is a precise statement of Band's value. The core insight (single LLM can't reliably reconcile across conflicting business rules, and you can't locate which step failed) is a real and important limitation addressed by role-based design. Solo submission with thorough validation. Business value: Cross-system financial reconciliation (CRM vs ERP vs Finance) is a universal enterprise finance pain point. Localised failure tracing (anomaly diagnosed to the specific agent and rule that identified it) satisfies compliance audit requirements that black-box outputs cannot. Originality: Root-Cause Agent diagnosing discrepancies using the domain business rules (payment terms, FX policies, invoicing schedules) retrieved by each system agent is a sophisticated domain-grounded approach. Three-tier verdict confidence scoring (normal/anomaly/watch) maps precisely to how finance controllers actually categorize findings."
Sanem Avcil
"I liked this project because it addresses a very practical enterprise reconciliation problem instead of treating AI as only a chatbot or reporting layer. The strongest part is the workflow design. The Router Agent directs the request to CRM, ERP, and Finance agents, which collect information from their respective systems. The Reconciliation Agent then identifies differences, and the Root-Cause Agent explains whether the discrepancy is expected, requires monitoring, or needs further action. This is a useful approach because reconciliation is not only about finding different numbers. The real business value comes from explaining why the numbers are different. For example, the project shows how a contract amount, invoice amount, and payment amount can differ because of installment timing, payment windows, bank fees, FX conversion, or a real data issue. I also liked the traceability design. The structured JSON handoffs, query ID isolation, and visible decision trail make the workflow easier to debug and audit. This is important in finance operations because users need to understand which system provided the data, what rule was applied, and why the final finding was classified as normal, watch, or anomaly. The project also shows good validation effort through multiple business scenarios, including installment payments, entity alias mismatch, bank fee adjustments, FX conversion, missing fields, invoice ID mismatch, and unexplained gaps. For future improvement, it would be useful to show stronger support for real production conditions such as late-arriving transactions, duplicate events, partial source-system failures, changing FX rates, larger data volumes, and human approval for high-risk reconciliation exceptions. It would also help to provide source-level lineage in the final report, showing the exact record, rule, and transformation that led to each conclusion. Overall, this is a strong multi-agent reconciliation solution with clear business value, practical architecture, and a good focus on traceability and root-cause analysis. "
Amit Singh
AVP
For Innovators & Creators