Most 'AI for finance' demos are a chat model wrapped around a single data call; the hard part isn't producing an answer, it's keeping an agent honest across a dozen dependent steps. Dexter treats that as the real problem — planning, iterative self-checking, and a full JSONL audit trail are first-class, on the premise that financial research breaks more from unbounded loops and unverifiable reasoning than from a weak model.
What Sets It Apart
- It drafts a research plan before acting, then pulls structured fundamentals (income statements, balance sheets, cash flows) from the Financial Datasets API instead of scraping — so numbers trace back to a source rather than being hallucinated.
- It reviews its own intermediate work and re-runs until confident, with explicit loop detection and execution caps — so it fails closed instead of spinning forever on an unanswerable query.
- Every tool call, argument, and reasoning step lands in JSONL scratchpads — so you can replay exactly how a conclusion was reached, which matters when the subject is money.
- It is model-agnostic (OpenAI by default, optional Anthropic, Google, XAI, Ollama) and reachable through a WhatsApp gateway — so you can swap providers or query it conversationally without rebuilding the stack.
Great Fit / Look Elsewhere
A great fit if you want to study how an autonomous research loop is wired together in TypeScript, or you need fundamentals-driven analysis with a verifiable paper trail. Look elsewhere if you expect trading signals or portfolio execution — the project is explicitly educational and informational, not financial advice, and its coverage is bounded by the Financial Datasets API and whatever provider keys you supply.
