Agentic AI / ML Systems · Fintech · production
MateFi
AI financial coach with safe-to-spend guidance.
Production personal-finance agent that connects accounts via Plaid, categorizes transactions, tracks goals, and coaches spending decisions—including a safe-to-spend number grounded in bills, subscriptions, and savings set-asides.
Problem
Bank balances lie. People overspend because apps show charts, not whether they can affo…
System
A retrieval- and rules-informed coaching agent over categorized transactions: safe-to-s…
Role
AI Architecture & Agent Orchestration
Status
production
Outcome
Safe-to-spend Day · week · month
My role
AI Architecture & Agent Orchestration
- AI Architecture
- Agent Orchestration
- Data Pipelines
- Retrieval
- Productionization
Period / 2024–Present
AI coach for everyday spending decisions
Plays once · tap to replay
Problem
Bank balances lie. People overspend because apps show charts, not whether they can afford a purchase after rent, bills, and goals.
System
A retrieval- and rules-informed coaching agent over categorized transactions: safe-to-spend computation, anomaly flags, goal decomposition, and an AI coach that suggests fixes without moving money until the user approves.
Architecture
Plaid-linked accounts sync transactions into categorization and subscription detection; a rules engine computes safe-to-spend and goals; an AI coach retrieves habit context to advise—never executing money movement.
- Bank Accounts → Transaction Sync
- Transaction Sync → Categorization
- Categorization → Safe-to-Spend Engine
- Categorization → Spending Memory RAG
- Safe-to-Spend Engine → Mobile App
- Goals Tracker → Safe-to-Spend Engine
- Spending Memory RAG → AI Coach
- Safe-to-Spend Engine → AI Coach
Execution path plays once on view · hover a node or tap the canvas to replay
System anatomy
Architecture
Bank sync → categorize → safe-to-spend → coach.
- Plaid connect
- Categorization
- Rules engine
- Coach agent
Engineering decisions
Decision
Safe-to-spend as the primary interface
Constraint
Balance-based UX drives overspending against upcoming obligations.
Approach
Compute spendable cash after rent, bills, subscriptions, and goal reserves for day/week/month horizons.
Result
A decision number users can act on instead of guessing from balances.
Decision
Coach that cannot move money
Constraint
Financial agents that auto-transfer create trust and liability risk.
Approach
The agent explains overspend math and suggested fixes; nothing moves until the user approves outside the coach.
Result
Informational coaching with clear non-custodial boundaries.
Decision
Plaid-mediated bank access
Constraint
Storing bank credentials is unacceptable for a consumer finance app.
Approach
Plaid handles authentication so MateFi never sees bank logins; data is encrypted in transit and at rest.
Result
Bank-level connect with privacy-first data handling.
Decision
Habit retrieval for personalized coaching
Constraint
Generic budgeting tips ignore a user's actual categories and goals.
Approach
The coach retrieves recent spending patterns and goal state before recommending category-level adjustments.
Result
Personalized guidance grounded in the user's transaction history.
Reliability
- No fund movement by the agentACTIVE
- Plaid-mediated authACTIVE
- Encryption in transit and at restACTIVE
- User-controlled unlink / data removalACTIVE
Evaluation
Outcome / Results
- Automatic categorization and subscription detection without manual tracking.
- Safe-to-spend guidance after obligations and goals—not raw bank balance.
- Privacy-first coaching: insights only; MateFi never moves user funds.
Stack
Product visuals
