AI FinSight

Inspiration

Let's face it: manual expense tracking is a chore. Most people abandon budgeting apps within a week because filling out endless forms, selecting categories, and typing exact amounts feels like homework. Living in a bilingual environment, we often mix languages or use colloquial terms (like spending on lunch or riding a rickshaw). We wanted to build a financial tracker that adapts to human behavior, not the other way around. We were inspired to create a frictionless financial companion where you can simply type or speak naturally in English, Bangla, or a mix of both and let AI handle the data entry instantly.

What it does

AI FinSight is an intelligent, bilingual financial sidekick that runs completely locally. Instead of manual form entry, users simply type or speak their expenses naturally. The application automatically: Extracts & Categorizes: Pulls structural items (amount, description, category) from conversational sentences. Visualizes Financial Health: Displays real-time metrics, budget rings, and interactive weekly/monthly transaction charts via a gorgeous glassmorphic dashboard. Coaches Your Finances:Generates localized AI reports that identify overspending habits and provide personalized, encouraging saving tips to keep you on budget.

How we built it

🛠️ How we built it

AI FinSight is built with a robust, modern, and completely local architecture: The Brain (AI Backend): Powered by FastAPI and Python. We integrated Ollama running Qwen2.5:3B locally on the machine. This ensures total data privacy (your financial logs never leave your device) and zero API token costs. The Face (Frontend): Built with Next.js 15, TypeScript, and TailwindCSS. We utilized shadcn/ui for UI components, Framer Motion for smooth glassmorphism animations, and Recharts to build the interactive financial dashboard. Voice & Data:We leveraged the native Browser Speech Recognition API for seamless voice-to-text inputs and used SQLAlchemy with SQLite for lightweight local storage.

Challenges we ran into

The biggest hurdle was building for a high-pressure hackathon environment where local LLMs can occasionally be slow or fail to return strict structured data.

Accomplishments that we're proud of

Accomplishments that we're proud of 100% Local Processing: We successfully built a full-featured fintech application that handles heavy tasks like speech-to-text processing, natural language extraction, and financial data analysis completely on the user's machine. *Zero Cloud Costs & Total Privacy:Because it runs entirely offline via Ollama, users get high-tier AI capabilities with absolute data privacy and exactly zero external API token fees. Flawless Bilingual Support:Achieving seamless, natural text parsing across English, Bangla, and mixed "Banglish" phrases feels like a massive win for localized accessibility.

What we learned

Small Model Optimization: We leveled up our prompt engineering skills, discovering how to force smaller 3B parameter models (like Qwen2.5:3B) to consistently output strict, deterministic JSON arrays without bleeding out conversational prose. Resilient Architecture Design: We learned the importance of engineering robust offline fallbacks. Creating a regex-and-keyword pipeline that takes over instantly if an LLM times out taught us how to design applications that never fail a live demo. Concurrent Resource Management: We gained deep insights into balancing web performance when multiple heavy processes such as browser voice recognition, local database mutations, and local LLM inference happen simultaneously on the client side.

What's next for AI FinSight

What's next for AI FinSight

• Local OCR Receipt Scanning: We plan to integrate a lightweight, local OCR engine so users can snap a photo of a physical receipt and instantly pull line items into their budget. Predictive Budgeting Insights:By implementing simple time-series data forecasting, the AI coach will be able to warn users mid-month if their current daily spending velocity will cause them to overshoot their budget. Cross-Device Encrypted Syncing: We want to introduce secure, peer-to-peer syncing so users can effortlessly track expenses across their phone and laptop without ever uploading data to a centralized cloud server.

Built With

• fastapi
• framer-motion
• next.js
• ollama
• python
• qwen2.5
• recharts
• speech
• sqlalchemy
• sqlite
• tailwindcss
• typescript
• web