Engineering Case Study: Project SiTepat

"The best technology disappears — the customer only sees trust." — Don Norman

SiTepat is a mobile application that provides independent access for customers of (Undisclosed Enterprise Bank) to monitor their financing status, savings, and installment schedules — without needing to visit a branch or contact an officer.

This article dissects the engineering challenges and architectural decisions behind developing a customer-facing application for the ultra-micro banking sector.

🏔️ 1. The Challenge

Conditions Before SiTepat

Before SiTepat existed, customers only had traditional ways to monitor their financing status:

• Physical installment books — manually recorded by field officers, prone to errors and loss.
• Contacting officers (Community Officer) — customers had to call or wait for officer visits to ask about balances and schedules.
• Visiting branches — consuming time and transportation costs, especially for customers in remote areas.
• No notifications — customers often forgot due dates, causing unintentional late payments.

Engineering Challenges

Building an app for banking customers presents challenges that are fundamentally different from internal apps (Tepati):

1. Security Surface Area: The app is exposed to the public via Play Store — the attack surface is much larger compared to internal apps distributed via MDM.
2. User Diversity: Target users are micro-customers — market traders, farmers, tailors — who might be using a smartphone for banking services for the first time.
3. Real-time Accuracy: Customers expect balances to update instantly after payment. A delay of just 5 seconds can trigger questions to the Call Center.
4. Legacy Integration: The bank's Core Banking System uses the ISO 8583 protocol (ATM/POS transaction standard from 1987) — needing to be bridged to modern REST APIs.
5. Scale Pattern: Traffic is bursty — massive spikes occur on due dates (the 1st, 5th, 10th, 15th of every month).
6. Device Fragmentation: Customers use Android devices with highly varied specs — from 1GB RAM to flagships, Android 5.0 to the latest.

🏗️ 2. Architecture Overview

Unlike Tepati, which has an offline-first architecture with message brokers, SiTepat uses an API Gateway pattern that abstracts the complexity of the Core Banking System from the mobile client.

graph TD
    A[Customer] -->|Open App| B[SiTepat Mobile App]
    B -->|HTTPS + Cert Pinning| C[API Gateway - Golang]
    C -->|Auth Check| D[Auth Service]
    D -->|Validate| E[OAuth 2.0 + MPIN/Biometric]
    C -->|Route| F[Account Service]
    C -->|Route| G[Financing Service]
    C -->|Route| H[Notification Service]
    F -->|ISO 8583| I[Core Banking System]
    G -->|ISO 8583| I
    H -->|FCM| B
    F -->|Cache| J[(Redis)]
    G -->|Cache| J
    I -->|Persist| K[(SQL Server)]
    C -->|Audit| L[(MongoDB)]

Key Architecture Decisions:

• API Gateway as a single entry point — all requests from the mobile app enter through one gateway handling auth, rate limiting, and request routing.
• ISO 8583 Bridge Service — a dedicated service that translates REST API calls into the ISO 8583 format understood by Core Banking, and vice versa.
• Redis Caching Layer — data that rarely changes (customer profiles, tenor details) is cached to reduce load on the Core Banking System.
• Circuit Breaker — if the Core Banking System is slow or down, the circuit breaker activates and displays data from the last cache, not an error.

🛠️ 3. Technology Stack

Backend (Built by me)

• Golang: API Gateway and microservices — chosen for low-latency and built-in concurrency to handle bursty traffic.
• Microsoft SQL Server: Primary database for transactional data.
• MongoDB: NoSQL database for storing massive customer audit logs.
• Redis: Distributed caching — TTL-based for balance data (30 seconds), longer TTL for profiles (5 minutes).
• Firebase Cloud Messaging (FCM): Push notification engine for payment reminders.
• OAuth 2.0 + JWT: Authentication and authorization framework.

Frontend (Mobile)

• React Native: Cross-platform app deployed to Android via Play Store.
• Secure Storage: Encrypted keychain to store tokens and MPIN hashes.
• Firebase Analytics & Crashlytics: Monitoring adoption and app stability.

Integration Layer

• ISO 8583 Adapter: Custom-built adapter to communicate with legacy Core Banking (Temenos/T24-based).
• Message Queue: For async processing of notification schedules and audit logging.

🔍 4. Feature Deep Dives

Feature #1: Core Banking Integration via ISO 8583 Bridge

Problem: The Core Banking System uses the ISO 8583 protocol — a fixed-length binary format designed in 1987 for ATM transactions. all balance, mutation, and financing data can only be accessed via this protocol. The mobile app needs a REST API.

Solution: I built an ISO 8583 Bridge Service — a Golang service acting as a translator between REST API (JSON) and ISO 8583 (binary).

graph LR
    A[Mobile App] -->|REST JSON| B[API Gateway]
    B -->|REST JSON| C[ISO 8583 Bridge]
    C -->|Pack to ISO 8583| D[Core Banking]
    D -->|ISO 8583 Response| C
    C -->|Unpack to JSON| B
    B -->|JSON Response| A

// ISO 8583 Bridge - Simplified Balance Inquiry
type BalanceInquiry struct {
    PAN         string `iso8583:"2"`   // Primary Account Number
    ProcessCode string `iso8583:"3"`   // 310000 = Balance Inquiry
    Amount      int64  `iso8583:"4"`
    STAN        string `iso8583:"11"`  // System Trace Audit Number
    LocalTime   string `iso8583:"12"`
    LocalDate   string `iso8583:"13"`
    MerchantID  string `iso8583:"42"`
}

func (s *BridgeService) GetBalance(ctx context.Context, accountID string) (*BalanceResponse, error) {
    // Build ISO 8583 message
    msg := &BalanceInquiry{
        PAN:         accountID,
        ProcessCode: "310000",
        STAN:        generateSTAN(),
        LocalTime:   time.Now().Format("150405"),
        LocalDate:   time.Now().Format("0102"),
    }

    packed, err := s.packer.Pack(msg)
    if err != nil {
        return nil, fmt.Errorf("failed to pack ISO message: %w", err)
    }

    // Send to Core Banking via TCP socket
    resp, err := s.conn.SendWithTimeout(ctx, packed, 10*time.Second)
    if err != nil {
        // Circuit breaker: fallback to cached balance
        cached, cacheErr := s.cache.Get(ctx, fmt.Sprintf("balance:%s", accountID))
        if cacheErr == nil {
            return cached.(*BalanceResponse), nil
        }
        return nil, fmt.Errorf("core banking unavailable: %w", err)
    }

    balance, err := s.packer.Unpack(resp)
    if err != nil {
        return nil, fmt.Errorf("failed to unpack response: %w", err)
    }

    // Cache the result
    s.cache.Set(ctx, fmt.Sprintf("balance:%s", accountID), balance, 30*time.Second)

    return balance, nil
}

Key Engineering Decisions:

• TCP Connection Pool: Maintain persistent TCP connections to Core Banking — avoid handshake overhead per request.
• Circuit Breaker Pattern: If Core Banking response time > 10 seconds or error rate > 50%, automatically fallback to cached data.
• STAN (System Trace Audit Number): Unique identifier per transaction for reconciliation and debugging.

Results:

Feature #2: Real-Time Financing Dashboard

Problem: Customers had no visibility over their financing status. Information like remaining ceiling, installment schedules, and payment history could only be seen in physical installment books — which were often lost or not updated.

Solution: A Financing Dashboard displaying comprehensive financing data in real-time.

// Financing Dashboard - Data Structure (React Native)
interface FinancingDetail {
  contractId: string;
  productName: string;
  disbursementDate: string;
  maturityDate: string;
  plafond: number;
  outstandingBalance: number;
  marginRate: number;
  tenor: number;
  installment: {
    monthly: number;
    nextDueDate: string;
    remainingInstallments: number;
    paidInstallments: number;
  };
  paymentHistory: PaymentRecord[];
  status: "active" | "completed" | "overdue" | "restructured";
}

interface PaymentRecord {
  date: string;
  amount: number;
  method: "auto_debit" | "cash_to_officer" | "transfer";
  status: "paid" | "late" | "partial";
  lateDays: number;
}

// Fetch with caching strategy
async function fetchFinancingDetail(
  contractId: string,
): Promise<FinancingDetail> {
  // Try local cache first (stale-while-revalidate pattern)
  const cached = await SecureStorage.get(`financing:${contractId}`);
  if (cached && !isStale(cached, 60_000)) {
    return cached.data;
  }

  try {
    const response = await api.get(`/financing/${contractId}`, {
      headers: { Authorization: `Bearer ${await getToken()}` },
    });
    await SecureStorage.set(`financing:${contractId}`, {
      data: response.data,
      timestamp: Date.now(),
    });
    return response.data;
  } catch (error) {
    // Offline fallback: return cached data even if stale
    if (cached) return cached.data;
    throw new FinancingFetchError(error);
  }
}

Key Design:

• Stale-While-Revalidate: Show data from cache first for instant UX, then update in the background. Customers don't need to see loading spinners.
• Payment Progress Ring: Visual indicator of what percentage of installments are complete — provides psychological motivation for customers.
• Overdue Alert: If there's a late installment, the financing card turns red with a clear call-to-action.

Results:

Feature #3: Smart Payment Reminder Engine

Problem: ~15% of late installments were caused by forgetting, not inability to pay. Previously, the SMS reminder system was ineffective as messages often went into spam folders and delivery couldn't be tracked.

Solution: Multi-channel smart reminder engine with configurable scheduling.

graph TD
    A[Scheduler Service] -->|Scan D-3| B{Due Date Check}
    B -->|D-3| C[Reminder: Gentle]
    B -->|D-1| D[Reminder: Urgent]
    B -->|D+0| E[Reminder: Due Today]
    B -->|D+1| F[Reminder: Overdue]
    C --> G[FCM Push Notification]
    D --> G
    E --> G
    F --> G
    G --> H[SiTepat App]
    A --> I[(PostgreSQL - Schedule DB)]
    G --> J[Notification Log]

// Smart Reminder Scheduler (Golang)
type ReminderSchedule struct {
    DaysBefore int
    Template   string
    Priority   string
    Channel    string
}

var reminderSchedules = []ReminderSchedule{
    {DaysBefore: 3, Template: "reminder_gentle", Priority: "normal", Channel: "push"},
    {DaysBefore: 1, Template: "reminder_urgent", Priority: "high", Channel: "push"},
    {DaysBefore: 0, Template: "reminder_dueday", Priority: "high", Channel: "push"},
    {DaysBefore: -1, Template: "reminder_overdue", Priority: "critical", Channel: "push+sms"},
}

func (s *ReminderService) ProcessDailyReminders(ctx context.Context) error {
    today := time.Now().Truncate(24 * time.Hour)

    for _, schedule := range reminderSchedules {
        targetDate := today.AddDate(0, 0, schedule.DaysBefore)

        customers, err := s.repo.FindByDueDate(ctx, targetDate)
        if err != nil {
            return fmt.Errorf("failed to fetch customers for %s: %w", targetDate, err)
        }

        for _, customer := range customers {
            // Skip if already paid
            if customer.IsPaidForCurrentPeriod() {
                continue
            }

            notification := Notification{
                CustomerID: customer.ID,
                DeviceToken: customer.FCMToken,
                Template:   schedule.Template,
                Priority:   schedule.Priority,
                Data: map[string]string{
                    "contract_id": customer.ContractID,
                    "amount":      formatCurrency(customer.InstallmentAmount),
                    "due_date":    targetDate.Format("2 January 2006"),
                },
            }

            if err := s.notifier.Send(ctx, notification); err != nil {
                s.logger.Error("failed to send reminder",
                    "customer_id", customer.ID,
                    "error", err,
                )
                // Don't fail entire batch for one customer
                continue
            }

            s.repo.LogNotification(ctx, notification)
        }

        s.logger.Info("reminders processed",
            "schedule", schedule.Template,
            "target_date", targetDate,
            "count", len(customers),
        )
    }
    return nil
}

Key Engineering:

• Batch Processing: Scheduler runs every morning at 06:00 WIB — processing all due customers in batch, not one by one.
• Idempotent: If scheduler runs twice (e.g., due to restart), customers won't receive duplicate notifications — deduplication based on customer_id + template + date.
• Overdue Escalation: If a customer is D+1 and hasn't paid, notification is also sent via SMS as a fallback — not just push notification.
• Skip Logic: Customers who have already paid are automatically skipped, avoiding annoying notifications.

Results:

Feature #4: Banking-Grade Security Layer

Problem: Banking apps are exposed to the public — every vulnerability could lead to theft of customer financial data. OJK and PBI (Bank Indonesia Regulation) require extremely high security standards.

Solution: Multi-layered security framework that I implemented on the backend and coordinated with the mobile team.

graph TD
    A[User Login] --> B{Biometric / MPIN}
    B -->|Valid| C[Generate JWT]
    C --> D[API Request]
    D --> E{Certificate Pinning Check}
    E -->|Valid Cert| F{Rate Limiter}
    F -->|Under Limit| G{JWT Validation}
    G -->|Valid Token| H{Device Integrity Check}
    H -->|Clean Device| I[Process Request]
    H -->|Rooted Device| J[Block + Alert]
    E -->|Invalid Cert| K[Reject - MITM Detected]
    F -->|Over Limit| L[429 Too Many Requests]
    G -->|Invalid Token| M[401 Unauthorized]

// Security Middleware Stack (Golang)
func SecurityMiddleware(next http.Handler) http.Handler {
    return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
        ctx := r.Context()

        // 1. Rate Limiting — per account, sliding window
        accountID := extractAccountID(r)
        if !rateLimiter.Allow(accountID, 30, time.Minute) {
            http.Error(w, "too many requests", http.StatusTooManyRequests)
            auditLog(ctx, "rate_limit_exceeded", accountID)
            return
        }

        // 2. JWT Validation with rotation support
        claims, err := validateJWT(r.Header.Get("Authorization"))
        if err != nil {
            http.Error(w, "unauthorized", http.StatusUnauthorized)
            auditLog(ctx, "invalid_token", accountID)
            return
        }

        // 3. Device Integrity Check
        deviceFingerprint := r.Header.Get("X-Device-Fingerprint")
        if isRootedDevice(deviceFingerprint) {
            http.Error(w, "device not supported", http.StatusForbidden)
            auditLog(ctx, "rooted_device_blocked", accountID)
            alertSecurityTeam(ctx, accountID, "rooted_device_detected")
            return
        }

        // 4. Session Binding — token must be from the same device
        if claims.DeviceID != deviceFingerprint {
            http.Error(w, "session invalid", http.StatusUnauthorized)
            auditLog(ctx, "device_mismatch", accountID)
            return
        }

        // 5. Request Signing Verification
        if !verifyRequestSignature(r) {
            http.Error(w, "invalid signature", http.StatusBadRequest)
            auditLog(ctx, "invalid_signature", accountID)
            return
        }

        ctx = context.WithValue(ctx, "claims", claims)
        next.ServeHTTP(w, r.WithContext(ctx))
    })
}

Security Layers Implemented:

Results:

Feature #5: Mutation History & Transaction Search

Problem: Customers often need transaction proof for their business needs — for example, proof of installment payment to apply for loans elsewhere, or ensuring deposits from officers have entered the account.

Solution: Searchable transaction history with filter and export capability.

// Transaction History Service (Golang)
type MutationFilter struct {
    AccountID  string     `json:"account_id"`
    StartDate  time.Time  `json:"start_date"`
    EndDate    time.Time  `json:"end_date"`
    Type       string     `json:"type"` // "credit", "debit", "all"
    MinAmount  *int64     `json:"min_amount,omitempty"`
    MaxAmount  *int64     `json:"max_amount,omitempty"`
    Page       int        `json:"page"`
    PageSize   int        `json:"page_size"`
}

func (s *MutationService) GetMutations(ctx context.Context, filter MutationFilter) (*MutationResponse, error) {
    // Validate date range — max 90 days per query (OJK requirement)
    if filter.EndDate.Sub(filter.StartDate) > 90*24*time.Hour {
        return nil, fmt.Errorf("date range exceeds 90 days maximum")
    }

    // Check cache for recent mutations (< 24 hours)
    cacheKey := fmt.Sprintf("mutations:%s:%s:%s",
        filter.AccountID,
        filter.StartDate.Format("20060102"),
        filter.EndDate.Format("20060102"),
    )

    if cached, err := s.cache.Get(ctx, cacheKey); err == nil {
        return cached.(*MutationResponse), nil
    }

    // Fetch from Core Banking via ISO 8583
    rawMutations, err := s.bridge.FetchMutations(ctx, filter.AccountID, filter.StartDate, filter.EndDate)
    if err != nil {
        return nil, fmt.Errorf("failed to fetch mutations: %w", err)
    }

    // Transform & enrich data
    mutations := make([]Mutation, 0, len(rawMutations))
    for _, raw := range rawMutations {
        mutations = append(mutations, Mutation{
            Date:        raw.TransactionDate,
            Description: enrichDescription(raw.Description),
            Amount:      raw.Amount,
            Type:        categorizeTransaction(raw.DebitCredit),
            Balance:     raw.RunningBalance,
            Reference:   raw.ReferenceNumber,
        })
    }

    // Apply client-side filters
    filtered := applyFilters(mutations, filter)

    // Paginate
    result := paginate(filtered, filter.Page, filter.PageSize)

    // Cache for 5 minutes
    s.cache.Set(ctx, cacheKey, result, 5*time.Minute)

    return result, nil
}

Results:

🚀 5. Overall Business Impact

The implementation of SiTepat changed how customers interact with the bank:

🏢 6. Post-Launch: Challenges & Evolution

6.1 Device Fragmentation Nightmare

One of the biggest post-launch challenges was Android device fragmentation used by customers. Unlike Tepati where devices were standardized by the bank, SiTepat runs on customers' personal devices:

Implemented Solutions:

• WorkManager for notification scheduling — more reliable than AlarmManager on Android 8+.
• Lite Mode: Reducing animations and image resolution for low-end devices.
• Adaptive Network: Detecting network quality and adjusting payload size.

6.2 Digital Adoption by Micro-Customers

A unique challenge not found in most banking apps: many micro-customers are using a banking application for the first time. Collaboration between engineering and product that we did:

• Onboarding Tutorial: Step-by-step interactive flow when opening the app for the first time.
• Simple Language: Not using technical banking terms. Example: "Loan Remaining" instead of "Outstanding Balance".
• Large Font Size: Minimum 16sp for readability, considering many customers are aged 40+.
• Community Officer-assisted registration: Field officers help customers register for the first time (bridging Tepati and SiTepat).

6.3 Play Store Compliance

Publishing a banking app on the Play Store requires additional compliance:

• Financial Services Policy: Must meet Google Play's financial services requirements.
• Target API Level: Mandatory update to the latest target API level every year — impacting permission models and background restrictions.
• Data Safety Section: Transparent declaration about data collection and encryption.

🎓 7. Lessons Learned

What Went Well

• Circuit breaker pattern saved UX when the Core Banking System underwent maintenance windows — customers could still view data from cache.
• Free Push Notifications replaced paid SMS — ROI was instantly felt in cost savings.
• Modular ISO 8583 Bridge made it easy to add new use cases (mutations, transfers, etc.) without changing Core Banking.
• Comprehensive Audit Logging helped the fraud investigation team resolve cases in hours, not days.
• Stale-while-revalidate caching made the UX feel instant even though Core Banking response times were variable.

What I'd Do Differently

• Start with GraphQL from the beginning for the Financing Dashboard — too many REST endpoints eventually returning similar data.
• Implement Feature Flags earlier — currently A/B testing and gradual rollouts are still quite manual.
• Invest in synthetic monitoring to proactively detect Core Banking latency degradation before customers feel it.
• gRPC between internal services — JSON serialization overhead was felt during traffic peaks.
• React Native New Architecture (Fabric + TurboModules) to reduce bridge overhead on low-end devices.

Remaining Technical Debt

• Migration to Jetpack Compose for native security modules.
• Implementing a more robust offline mode — currently fallback is only from cache, no local database layer like Tepati.
• Standardizing error taxonomy between ISO 8583 response codes and HTTP status codes.
• End-to-end encryption for in-app chat support.
• Optimizing app size — currently ~45MB, target < 30MB for customers with limited storage.

🎉 Conclusion

Project SiTepat is not just about building a banking app. It is proof that legacy technology (1980s) can coexist with modern expectations (2020s) if bridged with the right architecture.

As engineers, it's easy for us to get trapped chasing the latest technology hype. However, SiTepat taught me that the highest value of an engineer lies not in how "bleeding edge" the stack used is, but in the ability to solve real problems within existing constraints.