AuditX

Inspiration

My mom faced something similar at her company: employees spending more than they were supposed to, and it quietly hurt the business over time. There was no analyst team watching every receipt and timesheet, so unusual spending could pile up before anyone caught it. That stuck with me. Small finance teams need investigation capacity they do not have — something that finds patterns early, explains the evidence, and still lets a human decide what to do.

What it does

AuditX is an AI finance investigator for small businesses.

Employees submit expenses with receipts and weekly timesheets. Detectors look for duplicate receipts, abnormal spending against a person’s baseline, and timesheet conflicts (for example hours in one city and a receipt somewhere else the same day). High-confidence findings place a reversible hold on reimbursement and notify an admin. Other flags open a ranked case with evidence and a plain-language brief.

Admins review the queue, dig into an employee’s record, and accept or reverse holds. Employees can see their own score, findings, and why something is paused. The AI never decides guilt. Holds pause money, not people.

How we built it

We built a monorepo with parallel streams:

web/ — Next.js 15, Auth.js, Prisma, Tailwind, and shadcn for employee and admin UIs, with a fixtures mode so the demo runs unplugged
analysis/ — FastAPI for detectors, severity, baselines, scoring, and a synthetic company dataset
NVIDIA Nemotron for non-chat work: structured investigation briefs and judgment on ambiguous cases after rules catch the obvious
Local OCR (tesseract.js) for receipt prefill when needed

All demo data is synthetic — no real account numbers or financial records — to fit the Compound / PNC track rules.

Challenges we ran into

Venue Wi‑Fi and flaky APIs meant the demo had to work offline on fixtures first. We also had to keep Nemotron out of “chatbot mode” and into structured briefs and decisions. The hardest product call was holds versus punishments: we only pause reimbursement, with one-click reverse and an audit trail, so the system stays fair and defensible.

Accomplishments that we're proud of

A complete employee and admin loop that demos without the network. Reversible holds with live UI refresh when a hold clears. Plain-language findings with banned accusatory words in the employee UI. Detectors evaluated on synthetic seeds. A Private Capital design system that looks like real finance software, not a generic AI dashboard.

What we learned

When money and people’s records are involved, precision beats flashy recall. Employees must see the same reasons admins see, or the product feels unfair. Sponsor constraints (synthetic data, non-chat Nemotron use) made the idea sharper. And a personal story — watching someone you care about absorb slow financial leakage — is a better north star than a feature list.

What's next for AuditX

Wire Nemotron end-to-end on every ambiguous case with a clear eval, pull real receipt images from storage, optional voice readout of the top hold, and a careful path into real reimbursement systems — still never auto-deciding guilt.