Finance Assistance

All research & projects

SOFTWARE & DATA CASE STUDY

Finance Assistance

Finance Assistance is a local-first platform that converts heterogeneous financial records into traceable accounting and read-only analytics, with guarded machine-learning categorization and a planned path toward human-reviewed AI portfolio decision support.

Role

Independent developer: designed and implemented the evidence-first data architecture, exact financial contracts, ledger and reconciliation layers, typed query service, local dashboard, and guarded machine-learning categorization workflow.

Methods

API integration · canonical data modeling · reconciliation · deterministic categorization

System architecture

Layered architecture showing generic read-only financial inputs flowing through encrypted evidence stores and typed adapters into an exact canonical ledger, reconciliation, categorization, operations, typed queries, and Dashboard v2, with a dashed planned AI layer.

Finance Assistance separates read-only acquisition, encrypted evidence, exact canonical accounting, deterministic interpretation, operational state, and local presentation; future AI decision support is marked as planned.

Eight-stage pipeline showing acquisition, encrypted commit, normalization, ledger projection, reconciliation, categorization, guarded model evidence, and typed query validation.

Checkpointed evidence-first workflow from read-only acquisition to dashboard validation, with guarded model output treated as evidence rather than authority.

Contribution and implementation

  • Provider-independent contracts normalize accounts, transactions, balances, and positions while retaining exact values, source lineage, and immutable evidence.

  • A rebuildable balanced ledger and deterministic reconciliation preserve source history while separating accounting state, owner decisions, and operational health.

  • Read-only typed queries power a six-section local dashboard with bounded filters, exact serialization, data-quality indicators, and review surfaces.

  • The current machine-learning component is deliberately limited to guarded categorization; privacy filters, corroboration gates, deterministic precedence, and owner review keep model suggestions outside accounting authority.

Implementation: Python · SQLAlchemy · SQLite · AES-256-GCM

Results and evaluation

A checked-in synthetic regression report records 229 passing tests; the handoff did not rerun the suite.

A 14-case synthetic model evaluation recorded 41 of 42 exact decisions, handled five of six ambiguities, and exposed one overconfident error.

A checked-in recovery report records an encrypted backup and disposable restore drill across five local stores.

Interface

Synthetic local finance dashboard with six tabs, separate CAD and USD cards, confirmed-spending bars, data-quality states, and an owner-review attention panel.

Original synthetic representation of the six-section local dashboard; every account label and financial value is fictional.

Future direction

The roadmap adds explainable recurring analysis, forecasting, anomaly signals, investment accounting, and evidence-cited human-reviewed portfolio decision support after the data, recovery, and review gates mature.

Four-stage roadmap with implemented trusted data and guarded assistance followed by planned read-only intelligence and portfolio decision support, plus a boundary excluding autonomous trading and financial advice.

Implemented data and guarded-classification foundations lead to planned read-only intelligence and human-reviewed portfolio decision support.

Limitations and evidence maturity

  • The dashboard and guarded categorization workflow are local prototypes; live acquisition remains provider-dependent, and no current connection, synchronization, coverage, or provider-health claim is public.

  • The AI-assisted portfolio manager is a planned, human-reviewed decision-support direction. Autonomous trading, money movement, financial advice, production-readiness, and security-certification claims are outside the demonstrated scope.