AI-powered reconciliation
with a 3-tier engine

A 3-tier reconciliation engine powered by Gemini 3 Flash with agentic function calling, 8 investigation tools, and real-time streaming reasoning. Upload any bank statement — PDF, CSV, or scanned — and watch the AI match, learn, and explain.

Try Live Demo Sign In

14-day free trial · No credit card required · Cancel anytime

Dashboard showing extracted bank statement data

Agent Tools

3-Tier

AI Engine

40+

Currency Pairs

90-Day

Cash Flow Forecast

Sound familiar?

Bank statement processing is broken

Hours of manual data entry

Copying transactions from PDFs into spreadsheets. Typos. Missing entries. The same tedious work, month after month.

Unreadable statement formats

Every bank has a different layout. Scanned documents. Multi-page statements. Nothing matches your chart of accounts.

Month-end closing nightmares

Reconciliation takes days. Discrepancies appear at the worst time. Your team works overtime every month.

“I used to spend 15+ hours every month just entering bank data. Now it takes 15 minutes.”

— Sarah M., Bookkeeper

End-to-end processing

6-Stage AI Pipeline

Every document flows through six stages — each powered by a different Gemini 3 capability.

Core architecture

3-Tier Reconciliation Engine

Each transaction escalates only if the previous tier can't resolve it. Most match in Tier 1 — AI is reserved for genuinely complex cases.

Tier 1

Quick Scan

MethodRule-based, instant

ThinkingNo AI needed

BatchAll transactions

Speed< 1ms per match

Result93%+ auto-confirmed

Tier 2

AI Matching

MethodGemini 3 Flash

Thinkingthinking_level: LOW

BatchBatch 10, 4 concurrent

Speed~200ms per batch

Result60-92% confidence

Tier 3

Deep Investigation

MethodGemini 3 Flash

Thinkingthinking_level: HIGH

Batch1 transaction at a time

Speed~2s deep analysis

ResultComplex cases resolved

reconcile-engine.ts

// Tier 2: AI matching with dynamic thinking
const response = await ai.models.generateContent({
  model: "gemini-3-flash",
  config: {
    responseMimeType: "application/json",
    thinkingConfig: {
      thinkingLevel: ThinkingLevel.LOW,
    },
  },
});

Function calling

8-Tool Agentic System

Gemini 3 Flash calls real database tools in a manual function-calling loop — up to 10 iterations per investigation.

Gemini 3 Flash

searchTransactions

Fuzzy search by amount, date, keywords

searchInvoices

Vendor, amount, status filters

findInvoiceCombination

Subset-sum: invoices that add to payment

findPaymentCombination

Transactions that sum to invoice

getFXRate

Live + cached, 40+ currency pairs

searchCreditNotes

Credit notes explaining differences

getVendorHistory

Payment delay statistics

getLearnedPatterns

AI-learned vendor patterns

ai-agent.ts — function-calling loop

// Manual function-calling loop (up to 10 iterations)
while (response.candidates?.[0]?.content
  ?.parts?.some(p => p.functionCall) && iterations < 10) {
  const calls = parts.filter(p => p.functionCall);
  for (const call of calls) {
    await agentTools[call.name](...args);
  }
  iterations++;
}

Up to 10 iterations per investigation. Real Firestore queries. Streaming progress to the UI.

Live progress

Real-Time AI Reasoning

Watch the engine think. Every step streams to your browser via Firestore real-time listeners — not a spinner, but actual reasoning.

reconciliation_runs/progress

ProgressStream writes events to Firestore in real-time. Frontend renders via onSnapshot listener.

Scoring engine

5-Signal Match Scoring

Every potential match is scored across five independent signals. Total score out of 125 is converted to a confidence percentage.

Reference Match/40

38Found "INV-15998" in transaction

Amount Match/35

35ANG→USD exact after FX conversion

Identity Match/25

22"GlobalTech" fuzzy match 92%

Time Proximity/20

1415 days after invoice — within normal range

Context/5

4Unique amount, no duplicates

Total113/125 → 90% confidence

Built-in fee pattern detection

Stripe 2.9% + $0.30PayPal 2.9% + $0.30Wise 1%Square 2.6% + $0.10Card 3%

Continuous learning

It Learns From Every Match

Every confirmed match — manual or AI — updates vendor patterns in Firestore. The next time that vendor appears, matching is faster and more accurate.

GlobalTech Solutions

Vendor pattern — learned

AliasesGlobalTech, GT Solutions, GLOBALTECH INC

Confidence94% (23 matches)

Avg Payment Delay18 days (range: 12-25)

Keywordsstripe, globaltech, payment

ProcessorStripe (2.9% + $0.30)

How pattern memory works

Match Confirmed

Manual or AI-confirmed match triggers learning

Extract Keywords

Transaction description tokenized. Common words filtered. Top 15 kept.

Update Averages

Running averages for payment delay, amount range, and confidence — weighted by match count.

Better Next Match

Next time this vendor appears, patterns boost scoring. Levenshtein fuzzy matching (0.7 threshold) catches name variants.

// Levenshtein fuzzy matching
function stringSimilarity(a, b): number {
  const dist = levenshteinDistance(a, b);
  return 1 - (dist / Math.max(a.length, b.length));
}

// Threshold: 0.7 — catches "GlobalTech" ≈ "GLOBALTECH INC"

Anomaly reasoning

AI Fraud Detection & Anomaly Alerts

Statistical analysis runs on every transaction set. The AI doesn't just match — it watches for outliers, duplicate charges, slow payers, and FX exposure.

Statistical Outliers

high

Flags payments exceeding 2 standard deviations from the mean — catches unusual spikes automatically.

threshold = mean + 2 × stdDev

Slow Payer Detection

medium

Identifies vendors with 30+ day payment delays from AI-learned vendor_patterns. Alerts on cash flow risk.

typicalDelay.max > 30 days

FX Exposure Alerts

low

Detects cross-currency reconciliation matches where rate fluctuations could cause variance. Flags for review.

txn.currency ≠ doc.currency

Low Match Rate Warnings

high

When >20% of transactions remain unreconciled, the system raises a high-priority alert for investigation.

unmatchedRate > 20%

insights/page.tsx — anomaly detection

// Statistical outlier detection (2σ)
const mean = debits.reduce((a, b) => a + b) / debits.length;
const stdDev = Math.sqrt(variance);
const threshold = mean + 2 * stdDev;
const outliers = txns.filter(t => t.amount > threshold);

Anomalies computed client-side from real transaction data. No external fraud service — pure statistical AI.

Predictive intelligence

90-Day Cash Flow Forecast

Gemini 3 Flash with extended thinking (4,096-token budget) analyzes your transaction history, upcoming receivables, and vendor patterns to project cash flow with risk scoring.

Next 30 days

Projected Inflow$24,500

Projected Outflow$18,200

Net+$6,300

82%

30-60 days

Projected Inflow$19,800

Projected Outflow$21,100

Net-$1,300

68%

60-90 days

Projected Inflow$22,000

Projected Outflow$17,500

Net+$4,500

55%

What feeds the forecast

6 parallel Firestore queries

Accounts, transactions, invoices, bills, vendor_patterns, reconciliation_runs

Monthly cash flow aggregation

Last 6 months of credits vs debits, computed client-side

Upcoming receivables

All unpaid invoices with due dates and amounts

Upcoming payables

All unpaid bills with vendor and due date

AI match rate

Reconciliation quality score from latest run

Vendor pattern count

Number of AI-learned patterns improves forecast accuracy

Gemini 3 Configuration

financial-chat.ts — cashFlowForecast

const response = await ai.models.generateContent({
  model: "gemini-3-flash",
  config: {
    temperature: 0.3,
    responseMimeType: "application/json",
    thinkingConfig: {
      thinkingBudget: 4096,
    },
  },
});

Risk Identification

AI assigns severity (high/medium/low) and dollar impact to each risk factor

AI Recommendations

Actionable steps based on the analysis — chase invoices, reduce spending, build reserves

Natural language queries

Ask Your Books Anything

An agentic AI assistant with direct database access. Gemini 3 constructs its own Firestore queries — no hardcoded searches. Up to 10 rounds of tool calls per conversation.

queryFirestore

AI constructs arbitrary Firestore queries with filters, ordering, text search — across 11 collections

getFinancialOverview

Cross-collection aggregation: invoices, bills, accounts, reconciliation stats, vendor patterns

getCashFlowTrend

Monthly cash flow analysis with configurable lookback (up to 24 months)

Financial AI Assistant

Which vendors have unpaid invoices over $5,000?

Querying invoices: paymentStatus == unpaid, total > 5000

Found 4 results in invoices collection

Querying vendor_patterns for payment delay data

Complete — 2 queries, 2 rounds, 1.8s

I found 4 unpaid invoices over $5,000:

1. GlobalTech — $8,450 (15 days overdue)

2. Acme Corp — $6,200 (due in 5 days)

3. CloudSync Ltd — $5,890 (30 days overdue)

4. DataFlow Inc — $5,100 (due in 12 days)

CloudSync has a history of 30+ day delays per your vendor patterns.

Ask about your financial data...

Security whitelist: 11 Firestore collections. userId filter auto-applied. No raw database access exposed.

Loved by finance teams

See what our customers have to say about SmartInvoice

Saved 50+ hours/month

“I manage books for 23 clients. Before SmartInvoice, I spent 2-3 hours per client on bank reconciliation. Now? 10 minutes. I've literally 10x'd my capacity without hiring.”

Sarah Mitchell

Bookkeeper, 23 clients

12 bank formats, 1 tool

“The Gemini AI understood our weird foreign bank formats that no other tool could handle. We tried 4 competitors before finding SmartInvoice. Nothing else came close.”

Marcus Chen

CFO, Import/Export Co

99.8% accuracy achieved

“Our month-end close went from 5 painful days to literally 1 day. The accuracy is insane - I've found maybe 3 errors in 6 months of processing thousands of transactions.”

Linda Park

Controller, SaaS Startup

See the AI in action

Try the live demo — upload a bank statement and watch the 3-tier engine match, reason, and learn in real time. No signup required.

Try Live Demo Talk to Sales

No credit card required · 14-day free trial · Cancel anytime

AI-powered reconciliation
with a 3-tier engine

14-day free trial · No credit card required · Cancel anytime

// Tier 2: AI matching with dynamic thinking const response = await ai.models.generateContent({ model: "gemini-3-flash", config: { responseMimeType: "application/json", thinkingConfig: { thinkingLevel: ThinkingLevel.LOW, }, }, });

// Manual function-calling loop (up to 10 iterations) while (response.candidates?.[0]?.content ?.parts?.some(p => p.functionCall) && iterations < 10) { const calls = parts.filter(p => p.functionCall); for (const call of calls) { await agentTools[call.name](...args); } iterations++; }

// Levenshtein fuzzy matching function stringSimilarity(a, b): number { const dist = levenshteinDistance(a, b); return 1 - (dist / Math.max(a.length, b.length)); } // Threshold: 0.7 — catches "GlobalTech" ≈ "GLOBALTECH INC"

// Statistical outlier detection (2σ) const mean = debits.reduce((a, b) => a + b) / debits.length; const stdDev = Math.sqrt(variance); const threshold = mean + 2 * stdDev; const outliers = txns.filter(t => t.amount > threshold);

const response = await ai.models.generateContent({ model: "gemini-3-flash", config: { temperature: 0.3, responseMimeType: "application/json", thinkingConfig: { thinkingBudget: 4096, }, }, });

AI-powered reconciliationwith a 3-tier engine

Bank statement processing is broken

Hours of manual data entry

Unreadable statement formats

Month-end closing nightmares

6-Stage AI Pipeline

Upload

Vision Scan

Extract

Categorize

Reconcile

Learn

Upload

Vision Scan

Extract

Categorize

Reconcile

Learn

3-Tier Reconciliation Engine

Quick Scan

AI Matching

Deep Investigation

8-Tool Agentic System

Real-Time AI Reasoning

5-Signal Match Scoring

It Learns From Every Match

GlobalTech Solutions

How pattern memory works

Match Confirmed

Extract Keywords

Update Averages

Better Next Match

AI Fraud Detection & Anomaly Alerts

Statistical Outliers

Slow Payer Detection

FX Exposure Alerts

Low Match Rate Warnings

90-Day Cash Flow Forecast

What feeds the forecast

Gemini 3 Configuration

Ask Your Books Anything

Loved by finance teams

See the AI in action

AI-powered reconciliationwith a 3-tier engine

Bank statement processing is broken

Hours of manual data entry

Unreadable statement formats

Month-end closing nightmares

6-Stage AI Pipeline

Upload

Vision Scan

Extract

Categorize

Reconcile

Learn

Upload

Vision Scan

Extract

Categorize

Reconcile

Learn

3-Tier Reconciliation Engine

Quick Scan

AI Matching

Deep Investigation

8-Tool Agentic System

Real-Time AI Reasoning

5-Signal Match Scoring

It Learns From Every Match

GlobalTech Solutions

How pattern memory works

Match Confirmed

Extract Keywords

Update Averages

Better Next Match

AI Fraud Detection & Anomaly Alerts

Statistical Outliers

Slow Payer Detection

FX Exposure Alerts

Low Match Rate Warnings

AI-powered reconciliation
with a 3-tier engine

AI-powered reconciliation
with a 3-tier engine