FinRak

Inspiration

Managing personal finances today is broken. Tracking daily expenses feels like a chore, transactions are scattered across multiple digital wallets and bank apps, pending receivables have no clean home, and manual spreadsheets demand constant upkeep. Existing apps make it worse: cluttered dashboards, dropdown fatigue from endless category selects, and zero proactive intelligence. The core insight driving FinRak is simple: users naturally think in language, not forms. Nobody wakes up thinking in spreadsheet rows. They think: "Spent $220 on clothes yesterday." We built FinRak to meet users exactly there.

What it does

FinRak is an AI-powered personal finance assistant that lets you log, query, and manage your finances entirely through natural conversation. No forms, no dropdowns, no manual categorization.

AI-Powered Entry - Type or speak transactions in plain English. "Grabbed some food with coworkers for forty bucks" is instantly understood and logged.
Structured AI Parsing - Gemini AI automatically extracts the amount, category, transaction type, and temporal date from your message and writes a clean, schema-perfect record.
Diverse Transaction Tracking - Handles Income, Expenses, Pending Receivables, and Payments, all within a single unified ledger.
Auto Categorization - AI maps unstructured descriptions to standardized parent categories and sub-tags automatically.
Analytics Panel - Instant visual charts for spending trends, category breakdowns, and dynamic monthly pattern alerts, powered by Plotly.

How we built it

FinRak uses a five-layer pipeline architecture:

Frontend - Streamlit UI serves as the interactive web portal, lightweight, fast to deploy, and fully conversational in layout.

AI Core - Gemini AI (via Google Generative AI API) acts as the NLP semantic parsing engine. Each user message is sent to Gemini with a structured system prompt that instructs it to return a strict JSON schema containing amount, category, sub_tag, type (income/expense/receivable/payment), and a resolved date. Ambiguous temporal references like "yesterday" or "last Friday" are resolved to ISO dates by the model at inference time.

Processing Layer - Python + Pandas coordinate the middleware logic, validating the Gemini JSON response, normalizing categories against a master taxonomy, handling edge cases and null fields, and preparing the final record for persistence.

Storage - Google Sheets API with OAuth 2.0 functions as the serverless cloud database. Each transaction is appended as a row, making the ledger human-readable, exportable, and zero-maintenance. Google Sheets was chosen deliberately: it is free, cloud-hosted, and gives users direct visibility into their own data.

Analytics - Plotly renders interactive charts directly inside Streamlit, including expense breakdowns by category, monthly income vs. expense comparisons, and trend lines.

Challenges we ran into

Prompt engineering for reliable JSON output - Getting Gemini to consistently return valid, schema-perfect JSON (not prose, not markdown-wrapped JSON) required careful system prompt design with explicit output contracts and fallback validation logic.
Temporal date resolution - Phrases like "the other day" or "last week Thursday" are genuinely ambiguous. We had to instruct the model to make a best-effort ISO date resolution relative to the current date, and handle cases where it could not confidently resolve.
Google Sheets concurrency - Appending records from a stateless Streamlit session to a live Sheet without race conditions or row collisions required implementing sequential write controls using the Sheets API's append vs. update patterns.
Category normalization - Users describe the same expense in dozens of ways. Building a taxonomy that Gemini could map reliably, without over-categorizing into too many bins or under-categorizing into too few, took multiple iteration rounds.

Accomplishments that we're proud of

Zero-friction logging - A transaction can be logged in under 5 seconds with a single natural language sentence, compared to 6+ steps in traditional finance apps.
Fully serverless architecture - The entire stack runs on Streamlit Cloud + Google Sheets with no backend server, no database fees, and no DevOps overhead.
Robust semantic parsing - FinRak correctly handles genuinely ambiguous inputs like "grabbed some food with coworkers for forty bucks" and maps it to {amount: 40, category: Food & Dining, type: Expense} with high consistency.
End-to-end working prototype - All four phases of the development roadmap through the Analytics Dashboard (Phases 1 to 3) are fully functional in the submitted build.

What we learned

LLMs are remarkably capable structured data extractors when given precise system prompts with explicit output schemas. The key is treating the model as a parser, not a chatbot.
Google Sheets is a surprisingly powerful serverless database for lightweight applications: free, collaborative, version-controlled, and natively readable by non-technical users.
Friction is the enemy of habit formation - reducing logging to one sentence dramatically changes whether users will actually maintain their financial records long-term.
Prompt iteration is engineering work - writing, testing, and refining the Gemini system prompt was as rigorous as writing backend logic.

What's next for FinRak

Receipt OCR scanning - Camera-based capture of physical bills to automatically extract line-item transactions without any typing
Multi-currency support - Handle international transactions with automatic currency normalization
Contextual AI advisory - Proactive savings and investment goal recommendations based on spending history patterns
Push budget warnings - Trigger-based alerts when category spending approaches user-defined thresholds or when payment deadlines approach
React Native / PWA mobile app - A standalone mobile build for on-the-go logging, fully offline-capable
Public hosting release - Open FinRak to public users on a shared Streamlit Cloud deployment with individual Google Sheets backends per user