docs/PPT_STRUCTURE.md

BankBot AI — Presentation Structure (12 Slides)

---

Slide 1 — Title Slide

BankBot AI An AI-Native Financial Operating System

Subtitle: Production-grade · Real-time · Intelligent

Visual: Dark glassmorphism background, animated gradient orb, tech badges

---

Slide 2 — Problem Statement

Traditional banking apps are passive. BankBot is intelligent.

Problem	BankBot Solution
Banks show data, not insights	AI-powered financial analysis
No real-time fraud intelligence	Live anomaly detection scoring
No personalized coaching	Financial health score + AI briefing
No scenario planning	What-If Simulator with 36-month projection
No behavioral analysis	Spending heatmap + pattern detection

Visual: Side-by-side comparison screenshot

---

Slide 3 — Solution Overview

Your AI Financial Twin

5 core capabilities:

1. Real-time AI chat with full financial context
2. Predictive balance forecasting (conservative/expected/optimistic)
3. Automated fraud detection (4-factor scoring)
4. Behavioral spending analysis (heatmap, patterns)
5. Interactive what-if financial simulator

Visual: Dashboard screenshot

---

Slide 4 — System Architecture

Production-Grade Full-Stack Architecture

Next.js 14 Frontend
    ↕ HTTPS / WSS
Nginx (TLS + Rate Limiting)
    ↕
FastAPI Backend (33 routes)
    ↕              ↕              ↕
PostgreSQL 15   Redis 7      AI Engine
→ SQLite        → memory     OpenAI → Groq
  fallback        fallback   → Ollama → offline

Key point: Every layer has a fallback — the system never fully fails

---

Slide 5 — AI Intelligence Engine

4-Tier AI Fallback Chain

Priority 1: OpenAI GPT-4o-mini    (fastest, most capable)
    ↓ if unavailable
Priority 2: Groq llama-3.3-70b    (free tier, very fast)
    ↓ if unavailable
Priority 3: Local Ollama llama3   (fully offline)
    ↓ if unavailable
Priority 4: Rule-based engine     (always available)

Context injected per message:

• Live account balances · Transaction history
• Financial goals · Investment portfolio
• Behavioral patterns · Health score

Visual: Chat page with streaming animation

---

Slide 6 — Real-Time WebSocket Architecture

Streaming AI + Live Updates

Browser ──WS connect──► FastAPI
Browser ──{ type: "chat", message: "..." }──►
         ◄──{ type: "chat_start" }──
         ◄──{ type: "chat_chunk", content: "H" }──
         ◄──{ type: "chat_chunk", content: "er" }──
         ◄──{ type: "chat_end" }──
Browser ──{ type: "ping" }──► (every 25s)
         ◄──{ type: "pong" }──

Features:

• Character-by-character streaming
• Heartbeat every 25s
• Exponential backoff reconnect (1s → 2s → 4s → 30s max)
• HTTP fallback when WebSocket unavailable
• Prompt injection prevention (9 regex patterns)

---

Slide 7 — Financial Intelligence

AI-Powered Analytics

• Health Score — 100-point composite (6 dimensions)
• Spending Heatmap — weekly activity patterns
• Category Intelligence — AI insights per spending category
• Net Worth Timeline — balance trajectory
• Behavioral Analysis — late-night spending, weekend patterns
• Subscription Optimization — detect unused subscriptions

Visual: Analytics page screenshot

---

Slide 8 — Fraud Detection Algorithm

Real-Time Anomaly Scoring

Transaction → Score 4 factors:
  Amount spike    > 3.5x avg  → +40 pts
  Timing anomaly  11PM–4AM    → +25 pts
  Rapid-fire      < 3min gap  → +20 pts
  Duplicate       same+10min  → +30 pts

Score ≥ 30 → logged to fraud_logs
Score ≥ 50 → status: "flagged"
Score < 30 → status: "verified"

Visual: Fraud alert notification screenshot

---

Slide 9 — Performance & Caching

Cache-Aside Pattern with Auto-Fallback

Endpoint	Cold	Cached	TTL
Dashboard	65ms	10ms	2 min
AI Score	~2s	10ms	10 min
AI Briefing	~3s	10ms	1 hr
Transactions	18ms	—	—

Optimization: Query.with_entities() — column-only queries, no ORM hydration Result: 32x speedup (2.1s → 65ms)

Cache: Redis → in-memory fallback (automatic, zero config)

---

Slide 10 — Database Design

10-Table Normalized Schema

Core tables:

• users — profile, personality, AI settings
• accounts — checking/savings/investment
• transactions — 300+ with categories, tags, emotion labels
• goals — target amounts, AI-generated plans
• investments — portfolio with AI risk analysis
• subscriptions — with AI usage detection
• notifications — typed alerts (fraud/insight/warning)
• fraud_logs — risk scores, details, status
• ai_insights — cached AI-generated content
• analytics_snapshots — daily financial snapshots

Fallback: PostgreSQL → SQLite (automatic, same ORM code)

---

Slide 11 — Security & Observability

Production-Grade Engineering

Security:

• JWT (60min access + 7-day refresh rotation)
• bcrypt hashing (rounds=12, direct library)
• Rate limiting (120/min API, 10/min auth)
• Security headers (CSP, X-Frame-Options, etc.)
• Prompt injection prevention (9 patterns)
• CORS restricted to configured origins

Observability (GET /api/metrics):

• Request count, error rate, auth failures
• AI provider health (calls/errors/latency)
• Cache hit ratio
• Per-route timing (avg + max)
• Last 50 errors with timestamps

CI/CD: GitHub Actions (backend lint, frontend build, Docker smoke test)

---

Slide 12 — Demo + Conclusion

Live Demo

Demo flow (5 min):

1. Login → Dashboard (65ms load)
2. AI Chat → WebSocket streaming
3. What-If Simulator → live sliders
4. Analytics → heatmap + radar
5. System Status → live metrics

What makes this different:

• Not a prototype — deployable to production today
• Every feature backed by real data (301 transactions)
• AI that knows your finances, not generic advice
• Resilient architecture that never fully fails
• Full observability — you can see it working

Numbers: 33 routes · 14 pages · 10 DB tables · 65ms dashboard · 4-tier AI fallback

---

Presenter Notes

Opening line:

"Most banking apps show you data. BankBot understands it."

Closing line:

"This isn't a student project that happens to use an AI API. It's a production-grade system where AI is the core — every response is personalized, every insight is grounded in real data, and every layer has a fallback."

If asked about deployment:

"The frontend is on Vercel, the backend on Render with managed PostgreSQL and Redis. The whole stack can also run locally with a single command — run.bat on Windows."

If asked about the most impressive feature:

"The AI orchestration layer. It builds a personalized system prompt from the user's live database records, streams the response through whichever AI provider is available, and falls back gracefully through 4 levels. Very few projects implement this kind of resilient AI infrastructure."