WealthOS · Autonomous Wealth Operating System

Retail investors juggle stocks, mutual funds, ETFs, bonds, and NPS across brokers with no unified brain. Existing robo-advisors are black boxes: one model, one opinion, no accountability for being wrong. LLM trading bots are worse: they let a language model touch money directly.

WealthOS is architected around three laws: preserve the validated trading core, money always passes a gate, and everything is measured.

• A 21-agent analyst council produces isolated opinions (no groupthink), a devil's advocate agent attacks the consensus, and aggregation is calibrated: each agent's vote is weighted by its measured historical accuracy, not its confidence.
• No LLM output moves money. Every action passes a code-enforced veto gate and a five-ring risk system: position limits, portfolio limits, drawdown circuit breakers, spend budgets, and per-broker connection locks.
• Strategy changes face an evolution gauntlet with frozen statistical thresholds: a challenger must beat the incumbent across fixed gates before promotion. No silent drift.
• The platform is a modular monolith with ports-and-adapters seams: a framework-free core, broker adapters over Model Context Protocol, and a durable database-backed event bus. Sized for free-tier deployment today with documented seams for horizontal scale.
• Users converse with the system from Telegram: instructions in, trade and portfolio reports out.

Status: In development. The engineering substrate is locked before feature work: 15 frozen architectural invariants, 11 architecture decision records, and 15 subsystem PRDs that make the veto gate, five-ring risk system, and calibration loop enforceable rather than aspirational. The trading core derives from a previously validated system. I engineered a guardrail harness (policy hooks, frozen-path protection, specialist review agents) so implementation cannot violate those invariants. Build metrics publish here as phases ship.

Core JTBD. "When my money is spread across brokers and asset classes, help me act on it with institutional discipline, so I stop losing to inattention and emotion."

Value proposition. A council, not an oracle: 21 accountable analysts whose track records are measured, behind a gate that code controls. Trust is the product; autonomy is the feature.

MVP scope (MoSCoW). Must: portfolio aggregation, council analysis with calibrated output, gated paper execution, Telegram reports. Should: evolution gauntlet, web dashboard. Could: live broker execution, model selection. Won't (v1): social features, advice marketplace.

Key metric hypotheses (estimates, pre-launch). Activation: first portfolio connected within one session. Retention proxy: weekly Telegram report open rate. Trust metric: percentage of council recommendations users accept over time.

Compliance posture. Built against DPDP requirements and SEBI advisory principles as design constraints: consent-first data handling, no advice framing without registration, execution gated and logged.