onefi
personal finance command layer — planning, actuals, and kpi loops with ai-driven coaching across every account and entity.
what it is
onefi is a personal finance platform that consolidates every financial account, entity, and obligation into a single command layer. it connects to bank accounts via plaid, tracks transactions against budgets, monitors investment portfolios, and provides ai-driven coaching that's grounded in actual financial data — not generic advice.
the system manages finances across personal accounts, business entities, and investment vehicles as a unified picture. one person's financial life is complex — multiple bank accounts, a business, rental property income, investment accounts, tax obligations. onefi treats all of it as one system.
why it exists
personal finance tools solve the wrong problem. they track spending after the fact and show charts. what's missing is a system that understands your financial position across every dimension — income, obligations, investments, entities — and actively helps you make better decisions.
onefi is the financial layer of a personal operating system. it doesn't just record what happened. it models what should happen, compares actuals to plan, detects when things drift, and recommends course corrections before problems compound.
how it's built
scenario-driven architecture
the system is built on 25 behavioral scenarios that specify exactly how every feature should work. each scenario describes a real situation — "user connects a new bank account," "monthly budget exceeds target," "investment portfolio rebalance triggered" — with precise acceptance criteria.
this approach means the system's behavior is fully specified before implementation begins. the scenarios serve as both the design document and the test suite. when a scenario passes, the feature works. when it doesn't, the gap is visible.
the dark factory pattern
onefi uses the same "dark factory" approach as the other projects in this portfolio — a development model where the wall between code generation and code evaluation is the whole game.
the operator writes behavioral scenarios — narrative descriptions of what the system should do in specific situations. a separate evaluation process derives holdout tests from those scenarios. the builder agent reads the scenarios but never sees the evaluation criteria. same principle as train/test separation in machine learning: the agent that builds must never see the criteria it's evaluated against.
the result: 25 scenarios were written, evaluated, and validated in a single development phase. the eval suite runs on every change. regressions are caught before they ship.
financial data integration
the platform connects to financial institutions through plaid for real-time transaction data. historical data imports cover years of financial history, enabling trend analysis and pattern detection that wouldn't be possible with a few months of data.
each connected account feeds into a unified data model where transactions are categorized, tagged, and matched against budget targets. the system handles the complexity of multi-account, multi-entity financial management without requiring manual reconciliation.
current state
phase 1 is complete — all 25 scenarios written and validated. the platform architecture, data model, and evaluation framework are in place.
phase 2 focuses on eval holdout derivation and real-world data integration: personal account onboarding, historical transaction import, and portfolio tracking. the system will grow entity by entity — personal accounts first, then business entities, then investment vehicles.
the goal is a financial command layer that's as comprehensive and opinionated as the other systems in this portfolio.