$ cd ../projects ⌃O / -
elwin@sytadel:~$ ▋
mango
Agentic personal finance for LATAM: you talk to Claudio, you don't fill forms.
- Go
- Chi
- PostgreSQL
- Anthropic SDK
- Next.js
- Fly.io
# nota: Private project under active development.
## problem
Personal-finance apps die on friction: endless forms, categories you have to maintain and dashboards nobody looks at. Most people abandon them within weeks.
## approach
The principle is "agentic input, visual output". Everything comes in through a conversation with Claudio (an agent on Claude Sonnet 4.5) that reads and writes financial data via tool-calling — no forms, no visible CRUDs. To see the state, a read-only dashboard shows KPIs and trends over already-loaded data, without spending LLM tokens.
- "I spent 3000 on coffee with Jime" → logs the shared expense and categorizes it automatically.
- "I want to save for an iPhone in 6 months" → creates a goal with its savings plan.
- "How many dollars do I have today?" → checks the live blue-rate and computes it.
## architecture
A Go 1.25 + Chi API, PostgreSQL 16 on Neon, deployed on Fly.io (São Paulo region, low latency for LATAM).
- A custom tool-loop (Go + Anthropic SDK): Claudio proposes tool calls, the backend validates and executes them against real data.
- Dual model Sonnet 4.5 / Haiku 4.5 to balance cost and latency per task.
- sqlc for type-safe queries with no ORM; Goose migrations; Clerk auth.
## technical decisions
- Row-Level Security in PostgreSQL: per-user isolation is enforced at the database, not only at the application layer (defense in depth).
- Separating input (expensive, via LLM) from output (cheap, direct dashboard) to control token spend.
- An LLM-usage table to measure tokens and cost per interaction (observability with zerolog + Sentry).
- Tests with testcontainers: each run spins up a real, ephemeral Postgres.