Financial Intelligence Strategy Agent
An autonomous agent that turns live Canadian public data into decision-ready retail-banking strategy briefs, built on the Model Context Protocol, with responsible-AI guardrails as first-class requirements.
The business problem
Retail-banking strategy teams sit on top of rich public data (labour, rates, housing), but turning it into a current, defensible brief is slow manual work, and it goes stale the moment the next release lands. The temptation is to let an LLM "just summarize it," which is exactly where regulated advice and invented forecasts creep in.
This agent automates the synthesis while refusing to cross the lines that matter in a regulated domain.
The key decision
A BI agent "should" have a backend and a database. After verifying the source APIs were CORS-open, I declined both: fully client-side, $0/month, nothing to secure. And I treated the guardrails as requirements, not add-ons: no forecasts, no regulated advice, confidence-flagged outputs. In a regulated domain, what the agent won't do is the design.
Architecture
- Built on the Model Context Protocol (MCP), the agent calls live public-data adapters as tools.
- Live adapters for Statistics Canada, the Bank of Canada, and CMHC.
- A fully client-side architecture, chosen after verifying the source APIs are CORS-open, so it runs at $0/month with no backend to secure.
- GitHub Actions CI runs the deterministic acceptance suite on every change.
Governance & Responsible AI
Guardrails were designed as requirements, not afterthoughts:
- No rate forecasts: the agent reports and reasons over observed data, it doesn't predict.
- No legal or regulatory advice: a hard boundary for a regulated domain.
- Automatic Low-Confidence flagging on thin-evidence series.
- Source traceability on every figure: each number points back to its origin.
Evaluation approach
A deterministic acceptance suite (23/23 passing) pins the agent's behaviour: given known inputs, it must produce the expected, source-attributed output and honour every guardrail. Determinism is what makes an autonomous agent trustworthy enough to run unattended.
Lessons learned
Verifying CORS-open APIs first unlocked a radically simpler, backend-free design, a reminder that architecture decisions should follow the constraints, not the fashion. And in regulated domains, the guardrails are the product: what the agent refuses to do is as important as what it delivers.