LedgerMind, an agentic, grounded underwriter for revenue-based financing

LedgerMind is an agentic, grounded underwriter for revenue-based financing. Give it a small business's raw bank transactions and sales records, and it produces a structured, policy-compliant financing offer: a recommended advance, a repayment percentage and a written justification. A locally hosted LLM (Qwen 2.5 via Ollama) decides which analyses to run, such as revenue trends, volatility, cash-flow red flags and bank-to-sales reconciliation. It then writes the memo. The LLM never invents a number: every figure comes from deterministic Python tools, and a hand-rolled guardrail checks each number and policy rule against the computed ground truth. Any mismatch is rejected and regenerated. An evaluation harness runs the pipeline over synthetic merchants with known ground truth. It measures grounding faithfulness, policy compliance and consistency. The Streamlit interface lets you pick a merchant, watch the memo and offer generate, and view the evaluation report. All financial data stays on-device.