counterbill

What it does

Counterbill helps importers, clearing agents and small logistics teams investigate freight bills that conflict with their records. Users upload invoices, quotes, packing lists, measurements and correspondence. Six Mozaik agents extract cited facts, identify conflicting evidence, select comparisons, check contractual explanations and prepare an editable complaint with PDF and evidence-summary exports. When new evidence arrives, affected agents reassess the case, revise the findings and complaint, and preserve the earlier interpretation with an explanation of what changed. Application code performs the arithmetic. A billing difference is kept separate from any claim that a refund is owed. The system has completed local real-model investigations using Gemini’s free tier. Public examples disclose reconstructed documents and hypothetical terms. It does not send complaints or determine legal liability.

How the agents run concurrently

Counterbill uses @mozaik-ai/core with six agents: Evidence, Billing, Contract, Timeline, Case and Drafting. Evidence, Billing, Contract and Timeline start independent loops that run concurrently, sharing a case-specific RuntimeState containing documents, cited facts, calculations, contract interpretations, timeline entries and assessments. Coordination is event-driven: - Evidence publishes fact.recorded, triggering Billing and relevant Timeline work. - Billing selects comparisons; deterministic application code calculates the amounts. - Contract interpretations and calculation.completed events trigger Case reassessment. - assessment.updated triggers Drafting to revise the complaint. Adding evidence wakes affected agents. Revised findings supersede earlier versions, while complaint updates preserve user edits or propose changes for review. Persisted events and revision checks prevent duplicate processing and stale results. Each investigation runs in an isolated worker process. Recorded real-Gemini traces expose distinct loop IDs, timestamps and overlapping inference spans. Hosting processes one investigation at a time; agents within it still run concurrently.