Technical Architecture
Gemini 3.8 Flash with
agentic function calling
3-tier reconciliation with dynamic thinking levels, 8 investigation tools, streaming reasoning via Firestore, and pattern memory with Levenshtein fuzzy matching.
Full stack
System Architecture
Next.js 14 frontend, Firebase Cloud Functions backend, Gemini 3.8 Flash AI, and Firestore for real-time streaming.
- PDF scan (inlineData)
- CSV detection & parsing
- Template fingerprinting
- responseMimeType: JSON
- Tier 1: Rule-based scan
- Tier 2: thinking_level LOW
- Tier 3: thinking_level HIGH
- Pattern memory + learning
- queryFirestore tool
- getOverview / getCashFlow
- 90-day cash flow forecast
- Natural language queries
Gemini 3 API
How We Use Gemini 3
Four distinct Gemini 3 API features — Vision, Structured Output, Function Calling, and Thinking Levels — each serving a specific role.
Vision Processing
PDF pages sent as base64 to Gemini 3.8 Flash's multimodal input. Detects tables, headers, and transaction rows from any bank format.
inlineData: {
mimeType: "application/pdf",
data: base64EncodedPDF
}Structured Output
Guaranteed JSON responses — no regex parsing of AI text. Every extraction and reconciliation returns typed, validated data.
config: {
responseMimeType: "application/json",
// Gemini returns valid JSON directly
}Function Calling
8 tool declarations registered with Gemini. Manual calling loop processes tool calls until the AI is satisfied — up to 10 iterations.
tools: [{
functionDeclarations: [
searchTransactions,
findInvoiceCombination,
getFXRate, // ... 8 tools total
]
}]Thinking Levels
Dynamic thinking budget per tier. LOW for batch matching (speed). HIGH for complex investigations (depth). No thinking for rule-based.
thinkingConfig: {
thinkingLevel: ThinkingLevel.HIGH
// LOW for Tier 2, HIGH for Tier 3
}Deep dive
Reconciliation Engine
Five scoring signals, fee detection, FX conversion, combined payments, and pattern memory — all in one engine.
5-Signal Scoring Formula
Fee Pattern Detection
When a transaction amount doesn't match exactly, the engine checks if it matches after subtracting known processor fees.
Multi-Currency FX
- 40+ currency pairs with 3-tier fallback: Live API → Firestore cache → hardcoded rates
- Pegged currency detection — ANG, AWG, XCG treated as equivalent
- Cross-currency via USD pivot when direct rate unavailable
- FX tolerance: 2% for rounding, 5% for rate variance, 10% for fluctuation
Combined Payment Detection
When one transaction pays multiple invoices, the engine runs subset-sum with early termination to find document combinations that match the payment amount.
- Max 5 documents per combination (configurable)
- Groups by counterparty first (most likely), then cross-vendor
- Early termination after 3 matches found
- 2% tolerance for rounding differences
// Subset-sum with early termination
function findCombination(
docs, targetAmount, tolerance,
onFound, maxDocs = 5
) {
const recurse = (startIdx, current, sum) => {
if (found >= 3) return; // early termination
if (sum > target + tol) return; // prune
if (current.length >= 2 && diff <= tol)
onFound(current, sum);
};
}
Specifications
Technical Specifications
| AI Model | Gemini 3.8 Flash |
| Thinking Levels | LOW (batch) / HIGH (investigation) |
| Agent Tools | 8 function-calling tools |
| Max Iterations | 10 per investigation |
| Batch Size | 10 txns, 4 concurrent batches |
| FX Currencies | 40+ pairs, 3-tier fallback |
| Fee Patterns | Stripe, PayPal, Square, Wise, Card |
| Fuzzy Matching | Levenshtein distance, 0.7 threshold |
| Auto-Confirm | 93%+ confidence |
| Streaming | Real-time Firestore ProgressStream |
| Time Budget | 4 min/batch with graceful degradation |
| Combined Payments | Subset-sum, max 5 docs, early termination |
| Cash Flow Forecast | 90-day, 3 periods, 4096-token thinking budget |
| Anomaly Detection | 2σ outliers, slow payers, FX exposure, match rate |
| Chat Agent | Dynamic queryFirestore + 11 collections + 10 rounds |
| Chat Thinking | 2048-token budget, 90s timeout, 20-msg history |
Intelligence layer
Anomaly Detection & Cash Flow Forecast
Beyond reconciliation — statistical analysis, risk scoring, and predictive forecasting powered by Gemini 3 extended thinking.
Anomaly Detection
Statistical Outliers (2σ)
Debit amounts exceeding mean + 2 standard deviations flagged automatically. Runs on every transaction set.
Slow Payer Identification
Vendors with 30+ day delays detected from AI-learned vendor_patterns. Surfaces cash flow risks.
Low Match Rate Alerts
>20% unreconciled transactions triggers high-priority alert. Uses latest reconciliation_run data.
FX Exposure Detection
Cross-currency matches flagged for rate variance review. Detects when txn.currency ≠ doc.currency.
Spending Trend Analysis
Monthly debit comparison: >10% increase = trending up, >10% decrease = trending down.
// insights/page.tsx — computed on real data
const anomalies: { type, severity, description }[] = [];
// Statistical outlier detection
if (amount > mean + 2 * stdDev) anomalies.push(...);
// Slow payer from vendor_patterns
if (p.typicalDelay.max > 30) anomalies.push(...);
90-Day Cash Flow Forecast
Data Collection (6 parallel queries)
Gemini 3.8 Flash Configuration
Output Structure
Thinking Animation (7 steps)
✓ Analyzing historical transaction patterns...
✓ Scanning upcoming receivables and payables...
✓ Detecting seasonal spending trends...
✓ Computing projected inflows from unpaid invoices...
✓ Estimating recurring expense patterns...
✓ Assessing cash flow risk factors...
● Generating 90-day forecast with Gemini 3...
Agentic assistant
Financial Chat — “Ask Your Books”
An autonomous AI agent with direct Firestore access. Gemini 3 constructs its own queries from a full database schema — no hardcoded search functions.
Dynamic Query Architecture
queryFirestore — The Primary Tool
AI constructs arbitrary queries with filters, ordering, text search, and field selection. userId filter auto-applied for security.
Security Whitelist — 11 Collections
Agent Loop
Real-Time Progress Streaming
> ANALYZING QUERY
Processing: "Which vendors have unpaid invoices?"
> QUERYING DATABASE
Querying invoices: paymentStatus == unpaid
Found 12 results in invoices collection
Querying vendor_patterns: orderBy matchCount desc
Loaded financial overview across all collections
AI analyzing results (round 2)...
> GENERATING RESPONSE
Complete — 3 queries, 2 rounds, 2.1s
// Manual function-calling loop
while (response.functionCalls?.length > 0
&& iterations < 10) {
for (const fc of response.functionCalls) {
await writeProgressEvent(...);
result = await executeTool(fc.name, fc.args);
}
response = await ai.models.generateContent(...);
}
Export to any accounting system
Once your data is extracted, export it in the exact format your software needs.
Standard Formats
Ready-to-use templates for common accounting software.
Accounting Software
Pre-configured formats that import directly.
Custom Templates
Create your own export format with custom columns.
Enterprise Security
SOC 2 Type II
Certified compliant
256-bit Encryption
At rest and in transit
99.9% Uptime
SLA guaranteed
GDPR Compliant
Full data control
Ready to see it work?
Start your free 14-day trial today. No credit card required.