Aivinn — AI-Powered Financial Assistant

Aivinn is a production mobile app that helps users manage their personal finances through natural conversation. Instead of navigating menus or filling spreadsheets, users simply talk to an AI — logging transactions, setting goals, tracking budgets, and getting insights, all in one continuous chat experience.

The app is live on the App Store and Play Store, built entirely by me as the sole engineer.

The Problem

Most personal finance apps treat money as a collection of transactions. Aivinn treats it as a lived experience. The core idea: financial stress shapes everyday decisions, yet traditional tools still require users to open menus, categorize expenses manually, and remember to check their own progress.

Aivinn replaces that friction with conversation. A user types "dinner 20" and the AI logs it, categorizes it, and integrates it into their financial picture — instantly.

My Role

I designed and built Aivinn end-to-end as the sole engineer — from product architecture and database schema to mobile client, backend APIs, AI infrastructure, and deployment.

Core Features

Conversational finance — Users interact with their finances through chat. The AI understands natural language, processes receipts and PDF statements, categorizes transactions automatically, and detects recurring patterns like subscriptions and monthly bills.

Agentic memory — Unlike traditional chatbots, Aivinn remembers users across sessions. It stores structured summaries of goals, spending habits, and financial concerns, allowing the assistant to follow up naturally in later conversations.

Proactive insights — A background agent monitors transaction history and past conversations, surfacing gentle nudges when spending shifts, bills approach, or savings targets are at risk.

Multimodal input — Users can type, speak, or upload receipts, PDFs, and screenshots. The system extracts and processes the relevant information automatically.

Financial wellness score — A personalized metric reflecting income stability, spending behavior, savings progress, and upcoming commitments — giving users an immediate sense of their financial health.

Technical Highlights

Agentic Memory System — 97% Token Cost Reduction

The core challenge with chat AI: the more users talk, the more expensive every message becomes. LangGraph's default MemorySaver accumulates messages in a thread forever — a user with 200 messages sends all 200 with every new request.

I solved this by designing a session-scoped memory architecture:

• Every app open generates a fresh sessionId, creating a new LangGraph thread
• After each message, a background agent asynchronously compresses the last 30 messages into a 250-word summary stored in PostgreSQL
• On next session open, the summary is injected as context — the AI knows the user's full history without carrying the full history

Result: per-message token cost dropped from ~15,000 to ~450 tokens — a 97% reduction.

A dedicated case study covering the full implementation is available — architecture, database schema, cooldown logic, and session management.

SSE Streaming — 2.5x Faster Response Times

Blocking HTTP requests meant users stared at a blank screen for up to 9 seconds before seeing any response. I replaced this with SSE streaming using LangChain's streamMode and built a custom React Native SSE client to handle the stream on mobile.

Result: time-to-first-content improved from ~6–9s to ~2.5–4s (2.5x faster). Total response time reduced by ~60%.

A dedicated case study covering the SSE implementation is available.

Prompt Caching — 90% API Cost Reduction

Implemented Anthropic's prompt caching for the system prompt, which remains constant across requests. Combined with the session-based memory system which eliminated growing message history from context, this significantly reduced the total API spend.

Result: ~90% reduction in LLM API costs.

Performance

Drove the app's Lighthouse performance score from 64 → 96 through bundle optimization, lazy loading, and rendering improvements across the web client.

Ops Dashboard

Built an internal dashboard for the Aivinn team aggregating:

• Sentry — frontend and backend error tracking
• Grafana — server logs and performance metrics
• RevenueCat — subscription and payment data
• Database — user activity and wallet stats

All displayed in a unified shadcn UI interface, giving the team full visibility into app health and user activity without switching between tools.

Stack

Mobile:     React Native, Expo, TypeScript
Backend:    Node.js, Express, PostgreSQL, Prisma, Redis
AI:         LangChain, LangGraph, Anthropic Claude
Auth:       Clerk
Payments:   RevenueCat
Monitoring: Sentry, Grafana
Hosting:    VPS

Results