Fin Co-Pilot — 7-Agent AI Swarm Finance App

Problem

Personal finance apps are form-heavy, passive, and dumb. Users manually fill category dropdowns, struggle with charts, and get zero proactive guidance. I saw an opportunity to eliminate all of that — build a finance app where you just talk, and a swarm of specialized AI agents handles everything: parsing, categorizing, validating, learning your patterns, scanning receipts, tracking item-level spending, and coaching you in real-time.

My Role & Responsibilities

• Sole architect and developer — designed and shipped the entire system end-to-end across 15 milestones
• Designed the 7-agent AI swarm architecture on Gemini 2.5 Flash — each agent owns a specific cognitive task
• Built the premium conversational UI with animated typing indicators, dynamic quick-action chips, and contextual greetings
• Implemented receipt OCR via Gemini Vision — photograph any receipt and the AI extracts every line item automatically
• Built voice input (speech-to-text) for hands-free expense dictation
• Shipped a full Price Intelligence system — barcode/QR scanning, price history tracking, deal surfacing, purchase predictions, and store comparison
• Integrated biometric auth (Face ID / fingerprint), push notifications with intelligent thresholds, and PDF/CSV report export
• Migrated all agents to firebase_ai v3.4.0 SDK for production stability

Architecture

The intelligence layer is a fault-tolerant 7-agent swarm where each agent specializes in one task:

• Client Layer: Flutter app with premium chat UI, voice input, receipt camera, dashboard, and price intelligence
• AI Layer: 7-agent Gemini swarm (Orchestrator, Extractor, Validator, Context, Receipt OCR, Item Tracker, Pattern Learner)
• Backend Layer: Firebase Auth, Firestore, Cloud Functions, FCM, Crashlytics, Analytics
• Flow: natural language/voice/receipt input → orchestrated agent reasoning → validated transactions + insights → persistent storage + notifications

Flutter Client (Chat / Voice / Receipt / Dashboard)
                     ↓
        Gemini 2.5 Flash · 7-Agent Swarm
                     ↓
        Firebase (Auth / Firestore / Functions)
                     ↓
      Insights, alerts, reports, and coaching loops

Tech Stack

• Frontend: Flutter 3.x (Dart 3.5+) — premium Material Design 3 with dark/light mode
• State Management: Riverpod 2.x
• Navigation: GoRouter 15.x with custom page transitions
• AI / LLM: Google Gemini 2.5 Flash via firebase_ai — 7-agent swarm architecture
• Vision: Gemini Vision for receipt OCR and item extraction
• Voice: speech_to_text for hands-free input
• Backend: Firebase (Firestore, Auth, Storage, Cloud Functions, Crashlytics, Analytics)
• Charts: fl_chart 0.68 — spending trends, category breakdowns, budget gauges
• Notifications: Firebase Cloud Messaging + local notifications (budget alerts at 50%/80%/100%, daily summaries, price drops, coaching tips)
• Security: Biometric auth (local_auth), encrypted storage (flutter_secure_storage), Firebase Security Rules
• Scanning: mobile_scanner for barcode/QR code product lookup
• Export: PDF + CSV report generation with share functionality
• OTA Updates: Shorebird for instant over-the-air patches
• Animations: flutter_animate, Lottie, shimmer loading states

Platform Screenshots

Key Features Delivered

Conversational AI Copilot

The core interface — describe any transaction in plain English ("I spent $45 at Target on groceries") and the agent swarm handles intent classification, data extraction, validation, and confirmation. Smart follow-up questions ask only what's needed, one question at a time.

Price Intelligence System

A full price tracking engine: barcode/QR scanning for instant lookup, price history monitoring across stores, Deal of the Day surfacing, trending deals, wishlist with price drop alerts, store comparison rankings, and purchase predictions ("You usually buy milk every 4 days — need it soon?").

Receipt Scanning & Item-Level Tracking

Point your camera at any receipt — Gemini Vision OCR extracts every line item automatically. Not just "Groceries $47" but "Milk, eggs, bread, chicken" with individual item profiles and spending patterns.

Proactive Financial Coaching

The Pattern Learner agent monitors your behavior and surfaces timely coaching tips. Budget threshold alerts at 50%/80%/100%. Weekly AI-personalized financial wisdom. Daily spending summaries. Milestone achievements for financial wins.

Results & Impact

• 7-agent AI swarm — one of the most sophisticated multi-agent mobile architectures: Orchestrator, Extractor, Validator, Context Agent, Receipt OCR, Item Tracker, Pattern Learner
• 15 milestones shipped — from authentication to AI agents to price intelligence to final UI polish
• Zero-form expense tracking — the entire transaction flow is conversational
• Full product delivery — authentication, onboarding, dashboard, analytics, budgets, reports, notifications, price intelligence, coaching — not a prototype, a complete product
• SDK migration completed — upgraded all 7 agents to firebase_ai v3.4.0 for production stability

Challenges & Lessons Learned

• 7-agent coordination — defining strict responsibilities for each agent (extraction vs. validation vs. pattern learning) and building the Orchestrator to route and synthesize across all of them required careful prompt engineering and fallback chains
• Gemini Vision for receipts — OCR quality varies wildly across receipt formats; implemented multi-pass extraction with validation against expected totals
• Price Intelligence at scale — monitoring prices across stores with rate limiting, caching, and intelligent refresh scheduling
• Firebase Security Rules — UID-scoped rules with composite indexes across 15+ feature modules required careful planning to avoid query conflicts

How AI/Agents Were Used

This is an AI-native product at every layer. The app itself IS a multi-agent AI system — 7 specialized agents working in concert to handle every aspect of personal finance. Beyond the product, I used agentic VS Code workflows (GitHub Copilot Pro+ with custom agents) to accelerate development across all 15 milestones. The agents generated scaffold code, architecture proposals, and review suggestions — I validated, integrated, and shipped.