PredictCash

Análisis Predictivo de Flujo de Efectivo (30 Días)
Diagnóstico de Gastos y Liquidez
Simulacion de entradas y salidas
Historial
Plan de Ahorro
Plan de ahorro confirmacion

Inspiration

We were inspired by the harsh reality faced by Small and Medium-sized Enterprises (SMEs) in Mexico and Latin America. We noticed that the vast majority of businesses do not fail due to a lack of talent or sales, but rather due to poor management of their seasonal liquidity. Seasonal businesses suffer drastic revenue peaks followed by months of slow sales where fixed monthly expenses become a fatal trap. We wanted to create a tool that does not simply show past charts, but rather anticipates the financial future of the business and offers an active mechanism for capital protection and growth.

What it does

PredictCash is an autonomous and intelligent web platform designed to optimize SME treasury management. The application allows users to:

Predict cash flow over 30 days using a Machine Learning model and the fundamental daily treasury equation:

Final Balance_t = Initial Balance_t + Cash-In_t - Cash-Out_t

Check real-time banking data through integration with the Capital One (Nessie API) Sandbox environment.

Simulate live transactions (income and expenses) that instantly recalculate business metrics.

Protect capital with an Intelligent Savings Box that applies a $12.5\%$ annual yield via monthly compound interest:

A_n = (A_{n-1} + M) X (1 + {r/12})

Receive automated diagnoses and action plans powered by Gemini AI, featuring a fail-safe local fallback system.

How we built it

We built PredictCash by structuring a modular three-layer architecture:

Banking Data Layer (Sandbox Backend): Integration with the Capital One (Nessie) API using HTTP requests (GET and POST) to manage accounts, deposits, and purchases in real time.

Intelligence and Machine Learning Layer: Implementation of a predictive model using Scikit-Learn (Random Forest Regressor) to estimate seasonal inflows (Cash-In), backed by a conversational agent powered by Google Gemini AI for generating JSON-structured recommendations.

Presentation Layer (Frontend): Developed entirely in Streamlit (Python), with advanced custom CSS injection to achieve a high-level corporate neobanking visual identity (dark blue background, contrasting cards, and red action buttons).

Challenges we ran into

During the high-pressure hackathon development, we faced and overcame significant technical challenges:

Transaction Synchronization: Ensuring that deposit and purchase entries made in the live simulator immediately updated the Machine Learning model and recalibrated the liquidity curve without compromising interface fluidity.

Corporate Visual Consistency: Mastering advanced custom styling in Streamlit via CSS to eliminate default colors and deliver an impeccable aesthetic experience worthy of a professional financial application.

External Resilience: Designing a robust local fallback mechanism so that, in the event of any network latency or interruption with the Gemini AI API, the system maintains full functionality and delivers precise diagnoses without throwing screen errors.

Accomplishments that we're proud of

Successfully integrating a real sandbox banking environment (Capital One Nessie API) instead of relying on static or fictional mock data.

Successfully merging quantitative predictive analytics (Random Forest) with a viable business financial model (the compound-interest savings box).

Building a polished, responsive user interface that projects absolute confidence and corporate professionalism.

What we learned

We thoroughly understood the importance of combining quantitative algorithmic data prediction with a real business solution to solve critical economic problems for small business owners.

We learned to design modular software architecture under extreme pressure, prioritizing backend resilience and an exceptional user experience (UX/UI).

What's next for PredictCash

Expanded Banking Gateways: Connecting PredictCash with more Open Banking APIs to support multi-bank accounts.

Automated Alerts: Implementing proactive notification systems via SMS or email when the model detects an imminent liquidity risk in the coming days.

Tax Scenario Optimization: Adding tax estimation and withholding modules to offer comprehensive financial planning for micro-entrepreneurs.

Built With

Submitted to

HackMTY 2026

Created by

En el desarrollo de PredictCash, me encargué de toda la columna vertebral lógica y del Back-End del proyecto. Diseñé el motor de tesorería basado en la ecuación fundamental de flujo de caja y la lógica matemática de interés compuesto para la caja de ahorro inteligente con un rendimiento anual del 12.5%. Asimismo, implementé la integración asíncrona con el entorno de pruebas de la API de Capital One (Nessie) para simular operaciones bancarias en vivo, incorporé el modelo predictivo de Machine Learning con Random Forest (Scikit-Learn) para anticipar los flujos de efectivo, y programé la orquestación con Google Gemini AI junto con un sistema de respaldo local (fallback) para garantizar la máxima estabilidad operativa y precisión en los diagnósticos financieros.

FidelJnr Lozano Dominguez
I worked on the Front-end, giving it some design and color so that the user could distinguish the buttons. A color similar to the company's logo was chosen, aside from the fact that it's a dark color that doesn't affect the user's eye brightness as much. I feel that if there had been more time, it could have been complemented with more improvements.

I had already worked on a Dashboard which I made from scratch in PHP, HTML, MySQL, CSS, and JavaScript, with graphs, warnings, and design, which is linked with Ngrok, an ApiKey, to a Particle Photon 2 that is accompanied by MQ-7, MQ-135, BME680, and PMS5003 sensors, which are in charge of sending the particles in the environment in real time every 15 seconds, working together with Fidel Lozano and Brisa Villalobos, who were in charge of programming the Particle board and developing predictive models with Random Forest and Linear Regression.

LuisSammchez
Andres Lozano Dominguez
Brisa Guadalupe Jiménez Villalobos