tara - financial agent built using mastra
tara is an agentic financial research assistant that lets users ask natural-language questions about their transactions, spending trends, mutual fund returns, investment portfolio value and get grounded, database-backed answers instead of llm guesses.
built on the mastra agent framework, tara partitions its capabilities into three domain-specific tools (transactions, funds, and portfolio) so the model always resolves a query to the right context instead of picking from one giant function list. the system is designed around strict data grounding: dates are resolved relative to real system time before any query runs, and the agent is instructed to say data is missing rather than hallucinate figures.
tech stack
- agent framework: mastra
- llm: google gemini
- database: postgresql, drizzle orm
- runtime: node.js, typescript
- languages: typescript