Smart AI Finance Manager
Eemani Ananya Gimkalollu Hari Muncharla Munugala
Dasari Chaithanya,
Krishna, Prasad , Dept of Geethanjali, Dept Anurvesh Reddy,
Assistant Professor,
Dept of CSE (DS), CSE(DS), of CSE (DS), Sreyas Dept of CSE (DS),
Dept of CSE (DS),
Sreyas Institute of Sreyas Institute of Institute of Sreyas Institute of
Sreyas Institute of
Engineering and Engineering and Engineering and Engineering and
Engineering and
Technology, Technology, Technology, Technology,
Technology,
Telangana, India Telangana, India Telangana, India Telangana,
Telangana, India
ananyakrish124@gm hariprasadd02.05@ geethanjaligoud146 India
chaithanyareddybi ail.com gmail.com munugalaanurvesh
@gmail.com @gmail.com
reddy@gmail. com
Abstract— This paper introduces a smart personal finance At the backend, PostgreSQL serves as the primary data store,
management platform powered by artificial intelligence, designed accessed via Prisma ORM. This setup ensures efficient, secure,
to help users make informed financial decisions with ease. The and type-safe communication between the application and the
system offers comprehensive features such as support for multiple database, while also simplifying schema management and
bank accounts, budget planning, and real-time transaction
boosting development speed. For authentication and user
tracking. Users can assign default accounts, flag recurring
expenses, and receive timely alerts when spending exceeds management, the application integrates Clerk, a
predefined limits. Leveraging AI capabilities, the platform developer-friendly identity service that handles sign-ups, logins,
automates the categorization of expenses, extracts data from sessions, and multi-factor authentication. Clerk streamlines
scanned receipts, and generates actionable insights from uploaded security without compromising user experience, offering
bills. Additionally, it provides interactive charts and detailed ready-to-use UI components and API endpoints. The frontend
analytics to visualize spending patterns and promote better interface is built using Shadcn UI, a modern UI library
budgeting behavior. Monthly AI-generated reports deliver combined with Tailwind CSS for utility-first styling. This
personalized summaries and financial recommendations directly to results in a clean, responsive, and visually consistent design that
the user's inbox. Developed using modern tools like Next.js,
adapts smoothly across devices. Users visualize their financial
Tailwind CSS, Gemini AI, Prisma ORM, and Shadcn UI, the
platform ensures a secure, intuitive, and responsive experience data with Recharts, a library that powers dynamic bar graphs,
tailored for seamless daily use. summaries, and animated visualizations—making financial
analysis more intuitive.
Keywords—Next.js,Clerk(Authentication),Prisma(ORM),PostgreSQ
L(Database),Inngest(BackgroundJobs),Resend(TransactionalEmails) To further enhance automation, the app uses Inngest, an
,Shadcn/UI(ComponentLibrary),Recharts (DataVisualization),Gemini event-driven serverless platform that enables workflows like
AI (Data Categorization). sending weekly email summaries or alerts for unusual activity.
This intelligent automation minimizes manual tasks and keeps
users informed without effort. Together, these technologies form
a powerful ecosystem for managing finances in a smart, secure,
and user-centric way. This adds a layer of intelligent automation
I. INTRODUCTION that enhances the user experience and reduces manual
effort.Email notifications are dispatched through Resend, a
This project introduces a modern, full-stack web application modern API-first email platform known for its simplicity and
tailored for individuals and small businesses seeking greater high deliverability. Resend handles transactional email delivery
visibility and control over their finances. Built on the robust and efficiently, ensuring that users receive important updates on
scalable Next.js framework, the platform delivers a time, whether it's a weekly financial summary, a transaction
high-performance, developer-friendly experience designed for alert, or a threshold notification.
long-term maintainability and growth. The core functionality
centers around helping users track and manage their income and Overall, this application demonstrates a thoughtful integration of
expenses with ease. Through an intuitive dashboard and cutting-edge technologies across the stack—from data
interactive charts, users can view their financial habits, set management and authentication to UI design, automation, and
budgets, monitor savings, and gain actionable insights. The AI integration. It offers a compelling and intelligent solution for
system promotes financial awareness by allowing users to managing finances in a modern, digital-first world, balancing
categorize transactions, flag recurring costs, and receive usability, performance, and scalability.
proactive alerts when budgets are exceeded.
II. LITERATURE SURVEY and Resend enable event-driven automation and transactional
email communication. The frontend is built using Next.js for
[1] This project leverages a modern tech stack to speed and performance, Tailwind CSS for utility-first styling,
deliver a smarter and more flexible financial tracking and Shadcn UI for accessible, prebuilt components—together
solution than many existing tools. While popular platforms ensuring a user-friendly and responsive interface. The platform
like Mint and YNAB offer users features like budgeting and is designed with scalability, automation, and user empowerment
expense tracking, they often lack deep customization and as core principles.
automated workflows that can adapt to individual needs.
Our application aims to bridge that gap with a personalized, V. METHODOLOGY
automation-driven approach.
The development of the AI-enhanced personal finance
[2] At the heart of the system is Next.js, chosen for its
full-stack capabilities and fast rendering, which ensures a management platform follows a multi-faceted approach that
responsive and performant user experience. User draws upon principles from software engineering, artificial
authentication and session management are handled by intelligence, natural language processing (NLP),
Clerk, selected for its ease of integration, built-in UI human-computer interaction, and financial analytics. The
components, and support for secure login flows like methodology focuses on four core areas: (1) architecture design,
multi-factor authentication. (2) intelligent algorithm modeling, (3) data preprocessing and
transformation, and (4) interactive interface development.
[3] The user interface is crafted using Tailwind CSS Together, these elements support a system that is functional,
along with Shadcn UI, enabling the creation of sleek, secure, and scalable.
accessible, and responsive layouts. Financial data is made
visually understandable using Recharts, which supports A. System Architecture
dynamic, interactive graphs that help users analyze income,
spending, and budgeting trends over time.
[4] For backend automation, the app incorporates Inngest,
a powerful event-driven platform that handles workflows such
as sending routine reports or alerts based on user activity. To
ensure reliable and elegant communication, Resend is used
for dispatching transactional emails—like budget alerts and
monthly summaries.
III. OBJECTIVE
The primary objective of this project is to design and develop a
scalable, full-stack financial management web application
tailored for individuals and small businesses. The platform aims
to simplify the process of tracking income and expenditures,
while offering real-time visual insights into users' financial
behavior. Key features include intelligent transaction
categorization, budget planning, and automated email alerts—all Fig 4. System Architecture
designed to reduce manual effort and improve financial
awareness. By integrating tools like Clerk for secure user
authentication, Inngest and Resend for workflow automation and 1 Presentation Layer
email delivery, and Gemini AI for smart classification, the
application seeks to enhance user experience and enable The user interface is built using Next.js for fast performance and
informed financial decision-making. seamless routing, combined with Shadcn UI and Tailwind CSS
to create a responsive and modern design. This layer handles
IV. SCOPE everything users see and interact with — from logging in to
viewing dashboards, adding transactions, or managing accounts.
This project encompasses the development of a modern, The goal is to keep the experience intuitive while maintaining a
intelligent web-based finance tracker that enables users to clean, consistent design across devices..
securely manage, visualize, and gain insights into their financial
activities. The application supports multiple user accounts and 2 Authentication and Session Layer
offers features like expense tracking, automated receipt Security and access management are handled by Clerk, a
processing, and recurring transaction identification. The backend modern authentication service that offers pre-built UI
stack utilizes PostgreSQL for data persistence and Prisma for components and session handling tools. This layer governs
structured and type-safe database access. Google’s Gemini AI login/logout flows, session tracking, and route protection.
model enhances the system by automating transaction Clerk's components like SignedIn, SignedOut, and UserButton
categorization and extracting data from uploaded receipts. Clerk ensure that only authorized users can access sensitive areas. such
provides secure, streamlined user authentication, while Inngest as dashboards and transaction logs.
3 Application Logic Layer 3 Receipt Parsing
The core business logic of the application resides in this layer,
built using Next.js API routes and server actions. It processes all To minimize manual data entry and improve transaction
user commands—validating inputs, orchestrating workflows, accuracy, the platform includes a receipt processing pipeline.
and triggering AI-enhanced functions such as transaction This two-step process involves:
categorization, budget forecasting, and alert generation. The
modular architecture allows for easy maintenance, scalability, Text Extraction – Using OCR tools such as Tesseract OCR or
and testing of specific features without affecting the entire Google Vision API, the system converts receipt images into
system. readable text.
Information Structuring – After extracting raw text, Named
4 Smart AI Layer with Gemini Entity Recognition (NER) techniques and rule-based parsing
The platform uses Gemini, a large language model, to are applied to identify and extract key information such as:
intelligently process financial data. This includes automatically
categorizing transactions based on their descriptions, extracting C. Datasets
key information from uploaded receipts, and detecting patterns
in spending. Instead of manual data entry, the AI steps in to Public Data
reduce friction and offer intelligent insights — enabling users to
better understand and manage their finances. We made use of publicly available datasets that reflect
real-world financial behavior:
5 Automation and Email Notifications
For automation, the system integrates Inngest, which listens for ● UCI Bank Marketing Dataset: Useful for
specific events (like large transactions or weekly activity) and understanding financial patterns and segmentation.
triggers workflows accordingly. These workflows connect to ● Kaggle Synthetic Transaction Data: Offers realistic
Resend, a service that delivers dynamic, personalized emails to spending data across common categories like groceries,
users. Whether it’s a monthly summary, a budget alert, or a travel, and utilities.
spending report — users are kept informed in real-time without ● Plaid Sandbox API: Simulates actual banking
lifting a finger. transactions, allowing us to safely test features like
account linking and balance checks.
B. Algorithms
The platform incorporates a set of algorithms designed to These datasets were adapted to fit the structure of our
automate and enhance user financial interactions. application, including users, accounts, transactions, and budgets.
1 Transaction Categorization
Synthetic & Simulated Data
To help users gain meaningful insights into their spending
habits, each financial transaction is automatically classified into
Where public data was limited, we generated synthetic data
a predefined category such as "Food", "Transport", or "Utilities."
based on typical spending habits—such as monthly salaries,
This is treated as a supervised machine learning task,
recurring payments, and seasonal expenses. This helped in
specifically a text classification problem. Several approaches are
training and testing the models more effectively. In the future,
evaluated, including:
with user consent, anonymized app usage data can also be used
to improve prediction accuracy and personalization over time.
Multinomial Naive Bayes – suitable for handling
frequency-based text features.
D. Data Preprocessing
Support Vector Machines (SVMs) – known for strong
performance on high-dimensional data.
High-quality input data is vital for reliable AI predictions.
Transformer-based models like DistilBERT – fine-tuned for
Several preprocessing steps are employed:
financial text classification, providing contextual understanding
3.4.1 Cleaning and Normalization
of transaction metadata.
All text fields (e.g., merchant names) are converted to
lowercase and stripped of punctuation.
2 Budget Forecasting
Dates are normalized to ISO format (YYYY-MM-DD).
Missing fields are imputed using statistical measures or
Budget forecasting helps users anticipate future expenditures
domain defaults.
and manage their finances more proactively. This component is
3.4.2 Feature Extraction
built as a time series regression problem, where past spending
Temporal features such as day, month, weekday, and
behavior is used to predict future budget needs.
quarter are extracted from date fields.
Rolling averages, transaction counts, and total monthly
Linear Regression – for establishing baseline trends in user expenses are computed per user.
spending. TF-IDF and embedding vectors (via BERT or
Word2Vec) are created for text-based features.
Exponential Smoothing – for giving recent spending more 3.4.3 Dimensionality Reduction
weight in prediction. To enhance model performance and reduce computational load,
techniques such as PCA (Principal Component Analysis) and available or in limited amounts. This feature contributes to the
feature selection via mutual information are applied. user's financial protection by way of offering real-time active
3.4.4 Receipt Data Transformation monitoring of account activity.
Raw image inputs undergo preprocessing steps like grayscale
conversion, thresholding, and de-skewing before OCR.
Post-OCR, regex and NLP rules are used to extract and validate
VII.CONCLUSION
key-value pairs like merchant, amount, and date.
3.5 Evaluation Protocol
We run each algorithm on benchmark data sets using the This project provides a comprehensive and intelligent personal
standard measures: finance management system that bridges the gap between
Classification Models: Precision, Recall, Accuracy, F1-Score traditional budget tools and modern user needs. The goal was
Forecasting Models: RMSE, MAE, MAP not simply to track transactions or balances but to provide users
Anomaly Detection: ROC-AUC, Precision@k with a wiser, proactive, and personalized experience when
NLP/OCR Pipeline: Entity-level F1, character-level accuracy managing finances.
Cross-validation, holdout testing, and ablation studies are The system supports secure sign-up with Clerk authentication,
utilized for model validation linking or adding multiple accounts, labeling transactions, and
assigning monthly budgets with real-time alerts. The dashboard
features thoughtful visualizations, dynamic budget tracking, and
VI. RESULTS
modular account views, all delivered through a clean interface
constructed using Next.js, Tailwind CSS, and Shadcn UI.
The conceptualized AI-based personal finance platform yielded In addition, smart components like Gemini AI enable smart
promising theoretical outcomes in its various subsystems, scanning and organization of receipts, while Inngest and Resend
demonstrating the potential of intelligent automation in the arena operate in the background to send email summaries, reminders,
of money management. Machine learning models used for and notifications monthly. These integrations not only render the
transaction categorization were highly effective in system responsive but predictive as well, notifying users of
comprehending and organizing user expenditure patterns. overspending or revealing insights into spending habits.
Utilizing contextual embeddings from transformer-based With this system, users are provided with clarity, control, and
language models such as DistilBERT, the system was able to confidence over money. With architecture that is scalable and
categorize transactions with high accuracy into pre-defined design that is modular, this project sets the stage for the addition
categories such as food, travel, healthcare, and recreation. of additional features in the future like investment integration,
Compared to traditional frequency-based classifiers, the credit score analysis, or AI-based financial planning.
contextual model excelled at addressing merchant name This solution is a testament to the ways in which new
ambiguity and identifying subtle relationships between technologies—coupled with thoughtful UX and good backend
transaction descriptions even for short or noisy inputs. This logic—can reimagine the way individuals interact with and
translated into high precision and recall rates in categorization, manage their financial lives.
especially when applied to realistic, user-analog data generated
from publicly available financial datasets.
In budget forecasting, the system utilized time-series regression VIII.REFERENCES
models to provide predictive analysis of user expenditures.
Utilizing recurrent neural networks, specifically the Long
Short-Term Memory (LSTM) network, significantly enhanced [1] "The Role of Artificial Intelligence in Fintech"
the system's capability for forecasting future expenditure Authors: J. Gomber, R. J. Kauffman, C. Parker, B. W.
patterns by learning short-term trends and long-term temporal Weber
patterns in user expenditure histories. In comparison to more Journal: Journal of Management Information Systems
conventional models such as linear regression, the LSTM Year: 2018
architecture demonstrated lower prediction error and enhanced [2] "A Deep Learning-Based Approach for Automatic
resistance to seasonal and irregular expenditure patterns. Such Receipt Data Extraction"
forecasting capability is at the heart of the system's broader goal Authors: N. Smith, L. Cao
of promoting financial awareness, enabling users to establish Conference: IEEE International Conference on Document
realistic budget limits on the basis of their forecasted Analysis and Recognition (ICDAR)
consumption behavior. Year: 2021
Anomaly detection system, which served to initiate alerts [3] "Personal Financial Management Systems: An
regarding suspicious or out-of-pattern financial activity, Analysis of Functionality and Design"
employed unsupervised learning methods to identify outliers in Authors: J. D. Wells, J. S. Valacich, T. J. Hess
the transaction pattern. Through analysis of statistical anomalies Journal: Communications of the Association for
in the quantity, frequency, and categorical type of transactions, Information Systems
the system could identify unusual patterns that could be potential Year: 2011
instances of accidental over-spending or potential fraud. Having [4] "Automated Financial Document Analysis Using
Isolation Forests as the base detection algorithm permitted Machine Learning"
anomalous data points to be identified without prior specified Authors: M. Zhang, H. Liu
knowledge of what an "anomaly" was, making it suitable for Journal: ACM Transactions on Intelligent Systems and
adaptive real-world usage where labeled anomaly data isn't Technology
Year: 2019
[5]
"Financial Planning and AI: Challenges and
Opportunities"
Authors: T. M. Nguyen, P. W. Lewis
Journal: AI & Society
Year: 2020
In addition to automated processing, the service featured a
receipt-scanning module that enabled users to scan and log
physical purchases via the use of Optical Character Recognition
(OCR) and natural language processing combined. The module
utilized open-source OCR engines like Tesseract to recover text
data from scanned-in receipt images, which were then parsed
using specially constructed NLP pipelines in order to identify
primary financial entities like merchant name, total amount, and
transaction date. The system was highly effective in parsing
structured receipts, thereby extending the platform's ability to
retrieve data from beyond digital purchases and into paper-based
spending. While some limitations were noted in handling poorly
scanned and handwritten receipts, overall performance was
satisfactory for many uses.
