Inspiration
Almost everyone has experienced the frustration of looking at a monthly statement and noticing their bill is suddenly higher than expected. From sneaky telecom "regulatory cost recovery fees" and quiet 20% SaaS subscription price hikes to hidden delivery surcharges, companies quietly extract billions in junk fees every year. Most consumers and small businesses simply lack the time or patience to comb through hundreds of line items each month and draft dispute emails to customer support.
We built LedgerGhost to flip the script: an autonomous financial watchdog that does the tedious auditing for you, flags questionable charges with mathematical proof, and arms you with polite, effective dispute letters to demand your money back.
What it does
LedgerGhost is an autonomous audit agent designed to protect your wallet and business margins:
Universal Statement Ingestion: Ingests statements and invoices via file upload (CSV, statements) or instant 1-click sample data for rapid testing. Interactive Transaction Review: Extracts and categorizes every transaction, allowing users to inspect, verify, or adjust line items before auditing. Autonomous Fee & Anomaly Detection: Scans every line item against common patterns of deceptive billing—identifying stealth price creep, duplicate subscriptions, unannounced rate changes, and unitemized fees. Transparent Risk Breakdown: Provides clear risk classifications (Clear, Needs Review, High Risk) along with an explanation of why each charge is flagged and the exact amount you can recover. Instant Dispute Letter Drafter: Automatically generates professional, respectful, yet firm dispute letters citing the exact dates, amounts, and discrepancies, complete with one-click copy and download functionality.
How we built it
Architecture & Frontend: Built with React, TypeScript, and Tailwind CSS on Lovable, featuring full dark/light mode support, an intuitive financial dashboard, and responsive audit tables. Agent Intelligence with NVIDIA Nemotron on Nebius Token Factory: Nemotron Nano / Super: Handles fast, high-throughput parsing, transaction classification, and preliminary screening of line items to keep latency minimal. Nemotron 3 Ultra: Takes over for deep reasoning on flagged anomalies—calculating month-over-month percentage variances, cross-checking standard billing disclosures, and synthesizing customized, persuasive dispute letters tailored to the specific merchant. Dual-Mode Execution: Implemented a seamless toggle between a local instant demo mode (pre-loaded with realistic real-world billing anomalies for judges) and live Nebius Token Factory API integration.
Challenges we ran into
Balancing Model Latency & Reasoning Depth: Auditing dozens of transactions requires both fast initial screening and deep reasoning for disputes. We solved this by structuring a two-tier agent approach: letting lightweight models screen the volume while reserving Nemotron Ultra's heavier reasoning for contextual dispute generation. Avoiding False Positives: Legitimate price changes happen (such as prorated upgrades or tax adjustments). We tuned the agent's prompts and heuristics to highlight charges as “Potentially Unusual Charges” with clear rationale rather than making unsubstantiated accusations. Structuring Clean Output from Unstructured Text: Converting varied statement formats into reliable transaction objects with correct floating-point math required strict JSON schema enforcement in the prompt pipeline.
Accomplishments that we're proud of
Creating a truly closed-loop agent experience: the user doesn’t just get an analysis or a chart, they get a ready-to-send dispute letter that solves the problem. A frictionless 1-click demo that lets judges see the agent analyze, flag, and draft a dispute in under 30 seconds without needing to connect private bank accounts. A clean, modern UI that feels like a production-grade fintech utility rather than a rough hackathon prototype.
What we learned
How to leverage Nebius Token Factory to orchestrate specialized open-source models, selecting the right model size (Nano vs. Ultra) for cost-performance optimization. How critical tone is when drafting dispute communications: aggressive letters get ignored, but structured, polite, fact-based letters referencing exact amounts and dates get resolved quickly. The huge potential for autonomous agents in everyday consumer advocacy and financial defense.
What's next for LedgerGhost: Autonomous Discrepancy & Chargeback Hunter
Native PDF & OCR Statement Parsing: Integrating deep document OCR directly with multimodal Nemotron models to parse scanned bank receipts and multi-page PDFs. Email & Inbox Agent Integration: Allowing users to forward monthly bill confirmation emails to an autonomous LedgerGhost inbox for continuous background monitoring. Automated Dispute Dispatch: Integrating with email APIs so users can approve and send dispute emails to merchant support desks with a single click directly from the dashboard. Community Fee Database: Anonymously crowdsourcing recurring hidden fee patterns so the agent can alert users before they subscribe to known repeat offenders.
Built With
- ai
- api
- css
- nvidia
- typescript
