SUDEEP LALKA | FINANCE AGENT LAB
Independent portfolio project — Version 4

Explore 32 working finance prototypes and 96 editable synthetic scenarios across 7 working domains. Eight further agents are planned, including Credit & Lending and Audit & Controls. Follow a visual sample in seconds, open its working demo, or explore a connected investigation from original documents to a saved reviewer decision.

THE CONNECTED CASE
A supplier invoices seven units at $50 while the receipt supports six. The workbench identifies a $50 quantity exception. A $300 corrected invoice explicitly replaces the $350 original, preserving historical evidence. A reviewer records a recommendation against the exact evidence version.

FRAUD & PAYMENTS SPOTLIGHTS
Three Fraud & Risk prototypes explain payment-anomaly signals and payee-change evidence gaps. Six Payments prototypes cover payment status, settlement, refunds, FX and dispute evidence. A transport timeout does not establish failure. Editable cases retain missing or conflicting evidence and a clear human handoff; no retry, payment or dispute response is sent.

PAYABLES & CLOSE SPOTLIGHTS
Selectable visual stories lead into five Payables agents and seven Close agents. Supplier reconciliation preserves a supported historical revision explanation; PO analysis counts documented overlap once. Intercompany timing retains original cutoff balances. Journal support separates balance, amount coverage, accounting basis and review.

TREASURY & RECEIVABLES SPOTLIGHTS
Treasury has three working agents; Receivables has four, including multi-invoice cash application. Inspect current cash by currency while stale accounts remain excluded and the complete position stays unknown. Compare frozen forecasts with actuals, separating unmatched and future flows from paired metrics. Count a deduction and its representing credit once. Review collections priorities with distinct dispute and evidence holds.

REPORTING & FP&A SPOTLIGHT
Four agents cover budget variance, KPI consistency, management-reporting drafts and delivery-to-billing completeness. The KPI sample explains a 16 percentage-point definition gap between 44% aggregate margin and a 60% average of segment margins; it is not a performance change. The management sample explains a $3,400 expense variance through $2,000 volume, $1,200 unit cost and $200 fixed-cost contributors, with $0 ledger residual. Its source-linked deterministic draft leaves unsupported business causes as reviewer questions. The two new agents each have three editable scenarios.

WHAT TO INSPECT
- Completed visual stories with optional chapters and deeper editable cases.
- Defined source intake, immutable originals and exact source citations.
- Deterministic calculations, evidence gaps and downloadable reviewer packets.
- Evidence versioning, correction validation and stale-review protection.
- Saved reviewer notes and decisions in the connected workbench.
- Twelve connected-workbench regression fixtures with measured results.

AI AND REUSE STATUS
Standalone explanations and reporting narratives are original deterministic templates. The connected workbench's live model explanation integration is prepared but disabled until the provider is connected. No model accuracy or production savings are claimed. Five small attributed helper adaptations support selected checks; they do not integrate upstream finance platforms. OpenFPA contributes a pinned MIT ratio-of-sums helper for KPI comparison, not a treasury/forecast model or report-writing integration.

BOUNDARIES
All standalone scenarios use declared synthetic inputs and illustrative policies. Supplied evidence and approval flags do not authenticate documents or approvals. The connected invoice foundation supports defined CSV, searchable PDF and text templates, with no OCR or authenticity verification. Duplicate checks cover loaded records only. No payment, journal, customer message or report publication is executed. Owner review remains separate. The current Site remains private; personal-domain publishing is a separate step.