Neuraforage

BillBack: An agent that acts, and knows when it can't. Small businesses and kirana stores lose money to billing discrepancies because the real agreement lives in a WhatsApp chat or voice note, proof of delivery is a paper challan, and the invoice is a photo. With no enterprise ERP, checking bills requires cross-referencing three messy sources by hand, and owners rarely have the time to calculate compounding tax errors or negotiate settlements. BillBack is an autonomous agent that: 1. Multimodal Ingestion: Reads invoices, delivery challans, and WhatsApp rate screenshots using Meta's muse-spark-1.3 vision model. 2. Zero-Hallucination Audit: Overcharge calculations, line math (qty × rate), and 5% tax-on-overcharge are evaluated strictly by deterministic Python—eliminating any risk of LLM arithmetic hallucinations. 3. Multi-Round Autonomous Negotiation: Drafts itemized dispute emails with arithmetic proof and conducts up to 3 rounds of simulated negotiation. It enforces an 80% recovery floor: if vendor offers fall short, it safely escalates to human decision (Recovered: ₹0.00). 4. Refusal Guardrail: If given an unpriced document (e.g. a shipping label like hck.jpeg), the Extraction Guard detects missing rates and halts into an amber "needs_review" card—never certifying a bill it cannot read reliably.