FinSight

Inspiration

Budget apps feel reactive: they describe yesterday, not tomorrow.
Users juggle bank portals, spreadsheets, and manual categorization → low adherence.
Capital One / modern fintech patterns show how intelligent, secure, real-time systems should behave.
We asked: What if a personal finance tool behaved like a proactive co‑pilot—flagging drift early, predicting stress points, and coaching toward goals?

Unified dashboard: live balances, categorized transactions, cash flow & forecast.
AI/LLM powered transaction categorization with confidence scoring + fallback heuristics.
Predictive budgeting: spend drift & savings trajectory projections.
Anomaly & recommendation engine: surfaces overspend risk, unusual category spikes, and savings opportunities.
Scenario sandbox: experiment with adjustments to savings or discretionary spend.
Calendar integration: highlights upcoming recurring charges & predicted cash tension windows.
Consistent data layer across pages (no mismatched balances or stale numbers).
Built to evolve into a mobile-first push-driven companion (micro nudges & milestone alerts).

Frontend: Next.js 14 (App Router) + TypeScript + modular UI (shadcn/ui + custom chart components).
Backend/Data: Supabase (Postgres, Auth, Row-Level Security) for secure multi-tenant storage.
External Feed: Capital One Nessie API → normalized account + transaction tables.
Edge Functions: ML inference (spend prediction, enrichment) + LLM prompt orchestration.
AI Layer: GPT-3.5 for semantic categorization & narrative insight generation; lightweight statistical + rule hybrids for anomaly detection.
Real-Time: Supabase channels for live updates (near-instant dashboard refresh).
Utilities: Transaction normalization + formatting helpers to keep logic consistent across views.
Dev Workflow: Rapid iteration focused on correctness (percentage metrics, unified balances) before adding new surface features.

Challenge	Why it Mattered	Resolution
Savings rate showing 9000%	Broke trust in metrics	Audited pipeline; removed double scaling; standardized thresholds
Divergent transaction sources (dashboard vs transactions page)	Inconsistent UI	Centralized fetch + normalization layer
Confidence variability in AI categorization	Noisy recommendations	Introduced confidence gating + fallback deterministic mapping
Forecast volatility on sparse mock data	Unstable user experience	Added smoothing, variance clamps, minimum window guards
Maintaining cross-page balance consistency	Visual & cognitive mismatch	Shared dashboard data fetch logic reused in all feature pages
Time constraints (hackathon) vs. architectural cleanliness	Risk of brittle code	Early utility abstraction + modular UI primitives

Delivered a functioning end-to-end intelligent finance assistant (not a static mock).
Blended semantic AI + lightweight ML forecasting in under hackathon time constraints.
Established a secure, extensible architecture (RLS, edge functions, modular components).
Rapidly fixed core trust issues (metric accuracy + consistency) mid-build without derailment.
Built a future-facing roadmap (mobile-first evolution, predictive nudging) grounded in current architecture.
Clear differentiation: forward-looking anomaly & drift insights rather than passive logging.

Trust is earned through consistent numbers—fixing data drift early multiplies user confidence.
Hybrid AI (LLM + rules + statistics) hits a sweet spot for velocity + reliability.
Normalization utilities pay compounding dividends when expanding feature surface.
Forecasts without smoothing erode perceived intelligence—presentation matters.
Designing for mobile engagement from day one influences data model decisions (event granularity, subscription patterns).

Short-Term (Weeks):

Add PWA manifest + offline cache for recent transactions.
Push notification MVP (budget threshold, anomaly, savings milestone).
Refactor shared transaction utilities across all pages (remove duplication fully).

Mid-Term (1–2 Months):