FROM THE TEAM

Anatomy of one decision

The full path a Vestya recommendation travels — from your message, through a specialist's proposal, a rule engine, a risk review, and your approval, into a permanent record. Step by step, nothing skipped.

By the Vestya team · Published July 29, 2026

“AI-powered portfolio management” can mean almost anything, so here is ours with nothing skipped: the actual path from a message you type to a change in your portfolio, exactly as the system executes it.

1. Your message is classified, cheaply

When you write to Vera — “should I put some of this bonus to work?” — a small, fast model does one narrow job: deciding what kind of request this is. Chitchat, research, stocks, crypto, or a confirmation of something already proposed. No trading logic lives here; it is a router.

2. A specialist writes a concrete proposal

Research questions go to Rumi, who reads and summarises but is structurally unable to propose trades. Actionable ideas go to the specialist who owns the asset class — Sage for stocks, Cleo for crypto. A proposal is not vibes; it is a ticker, a side, and a dollar amount, produced with your goals, limits, holdings and recent history in context.

3. Deterministic rules screen it before any review

The proposal hits the rule engine — position caps, crypto ceiling, cash floor, leverage ban, wash-sale window, BTC/ETH-only. This is plain code. A proposal that fails is dead on arrival, the rejection is recorded with the rule that fired, and no model can overrule it.

4. Roan reviews what survives

A separate risk agent — running on a stronger model than the proposers, because judgment is the expensive part — reads the proposal against your risk posture. Approve, veto, or amend downward. Amendments can only shrink a trade; the system enforces that arithmetic in code.

5. Consequence decides who signs

Below your limits, in paper mode, small crypto trades fill automatically. Everything at $5,000 or above, every stock trade, and everything in real-money mode becomes a card with the reasoning and a yes/no. You are the last gate for anything that matters.

6. Execution is recorded by code, not narrated by a model

Fills are written to the ledger and the confirmation you read is rendered from those rows — the conversational model does not get to describe your account activity from imagination. In paper mode the fill uses real market prices against your practice balance.

7. One permanent record ties it together

Proposal, rule results, risk review, your decision, outcome — one immutable record, stamped with the exact model version and instruction fingerprint of every agent involved. That record is what “ask why on anything” is backed by.

The shape of the whole thing

Notice what the pipeline never does: it never lets one model research, decide, size, approve, and execute in a single unreviewed pass. Every step is either deterministic, adversarially reviewed, or signed by you. That is the entire design philosophy in one sentence — AI proposes, code constrains, you dispose.

Not investment advice. Investing involves risk, including loss of principal. Paper-trading results are hypothetical and for educational purposes only.

Meet the team · How it works, the short version