SettleSight

Inspiration

For freelancers and merchants collecting payments via PayPal, month-end bookkeeping involves far more than matching invoice totals against deposits. Processing fees (~2.9% + $0.30) cause balance variances, while partial installments, combined payments, and omitted invoice numbers create confusion. Debited refunds (T1107) and duplicate payments complicate ledgers further.

Sellers often spend hours auditing spreadsheets manually. We built SettleSight: a reconciliation desk combining deterministic mathematical checks with NVIDIA Nemotron on Nebius Token Factory to autonomously audit exceptions, evaluate fee math, and triage discrepancies for human approval.

(Note: SettleSight uses synthetic demo data and PayPal sandbox records; all metrics describe test scenarios rather than production revenue or private customer metrics.)

What it does

SettleSight delivers an autonomous reconciliation desk with an AI auditor powered by Nebius Token Factory:

• Normalizes & Auto-Matches: Converts invoice and transaction feeds into integer cents, auto-matching clean pairs with 1.0 confidence.
• 7 Reconciliation Heuristics: Flags fee deductions within tolerance, combined payments, partial installments, email matches, refund debits, and duplicates.
• Nemotron AI Triage Mode: An interactive toggle activates NVIDIA Nemotron to inspect exceptions, evaluate fee math, and output a prioritized queue with risk ratings (Critical, High, Medium).
• Deterministic Math Governance: Invariant validator (validateAllocation) strictly checks all proposals. If the LLM proposes an over-allocation, the engine flags an ❌ Invariant Rejection. AI advises; code validates.
• Two-Turn Function Calling: Nemotron executes structured tools (list_exceptions, explain_match, draft_reminder, triage_exceptions) with visible traces.
• Confirmation Gate: Approvals update local state. PayPal sandbox payment recording requires explicit confirmation (confirmedByUser: true).

How we built it

SettleSight combines React 19, TypeScript, Vite, Node.js, and AG Studio 3.0:

• Nebius Token Factory: Secure Node.js proxy (server/llm.mjs) proxies requests to Token Factory's OpenAI-compatible endpoint. API keys remain server-side.
• Runtime Model Discovery: Backend queries GET /v1/models at runtime to discover and cache active NVIDIA Nemotron models. It selects nvidia/nemotron-3-super-120b-a12b for complex triage and nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B for fast tool calls, failing cleanly if none exist.
• Deterministic Math Engine: Developed integer-cent matching rules and an invariant allocation validator in ESM JavaScript (src/match/).
• Submission Window Milestones (Aug 26 – Oct 30, 2026):
  • 2026-10-09: Dashboard scaffold (451750d), PayPal sandbox client (d894aae), and agent harness (87dc739).
  • 2026-10-10: Matching engine (9169af6), Node 22 CI (501dd33), and automated video pipeline (32cdd85).
  • 2026-10-11: Nebius Token Factory provider (7bfa495), live Nemotron triage and PayPal write-back (f012ea5), hybrid mode (c6aad99), and video fixes (927b4cc, 6920a47).

Challenges we ran into

• Undocumented Nemotron Model IDs: Token Factory docs featured Llama and DeepSeek, omitting exact Nemotron IDs. We implemented dynamic discovery against GET /v1/models to resolve nvidia/nemotron-3-super-120b-a12b and nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B.
• Model Latency Tradeoffs: Nemotron Super took ~14.7s for 8k-token triage (fine for queue ranking), while Nemotron Nano completed fast tool turns in ~2.1s. Splitting tasks by model size solved this latency hurdle.
• Vite Watcher & Edge Capture Interference: Headless Edge recording triggered Vite watcher reload loops. We isolated Edge's user profile in submission/video/ for clean automated captures.
• UTF-8 Title Card Encoding: FFmpeg drawtext corrupted UTF-8 punctuation (→, ·) on Windows. We resolved this by rendering cards as browser screenshots.
• Honesty Fixes: We aligned demo narration to clearly distinguish synthetic demo data from sandbox records and verified fee tolerance logic.

Accomplishments that we're proud of

• Live Nebius Inference Verification: Verified real Token Factory inference on 2026-10-11: Nemotron Super evaluated 23 candidates (8,312 tokens), with 15 passing mathematical validation and 5 invalid allocations correctly rejected.
• PayPal Sandbox Write-Back: Authenticated against PayPal sandbox with 30 test invoices, executing a confirmed write-back on invoice SS-DEMO-028 ($25.00) that transitioned to PARTIALLY_PAID (EXTR-18338036U4388992R).
• 80 Passing Automated Tests: Node native test runner passes all 80 tests across 4 suites covering math rules, API methods, and security guards.
• Hybrid Demo Presentation: Shipped an instant GitHub Pages demo alongside a hybrid mode displaying "Demo data · Nemotron Live".

What we learned

• How easily OpenAI-compatible apps adopt Nebius Token Factory with zero custom SDKs—just updating the base URL and Bearer token.
• Pairing generative AI with deterministic code: Nemotron provides flexible reasoning, while invariant checks protect financial ledgers.
• Optimizing multi-model LLM architectures by pairing lightweight models (Nemotron Nano) for tool calling with larger models (Nemotron Super) for triage.

What's next for SettleSight

• Multi-Currency FX Triage: Applying Nemotron reasoning to cross-currency exchange rate adjustments.
• Real-Time Webhooks: Event-driven triage triggered by PayPal webhook notifications.
• Accounting Sync: Automated reconciliation export to cloud accounting platforms like Xero and QuickBooks.

Built With

• ag-studio
• ffmpeg
• javascript
• nebius-token-factory
• node.js
• nvidia-nemotron
• paypal-sandbox
• react
• typescript
• vite