SlothBudget Pro：以 Gemma 3、TabPFN 和 Render 打造的自主離線理財智能體

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

SlothBudget Pro: An Autonomous Offline Financial Agent Powered by Gemma 3, TabPFN, and Render
What I Built
Manually organizing bank statements, UPI transaction histories, and raw receipt text logs is a tedious, error-prone chore that eats hours every month. SlothBudget Pro completely automates financial housekeeping in the background.

You simply drop raw receipt text files or transaction logs into a local folder (or upload them directly through the cloud web dashboard), and the autonomous agent handles unstructured text extraction, tabular machine learning classification, and ledger archiving—requiring zero manual data entry.

Demo
Live Cloud Dashboard (Render): SlothBudget Pro on Render ()

Video Demo / Screenshots:

Code
You can explore the complete open-source codebase on GitHub:
GitHub Repository: https://github.com/niveena01/SlothBudgetPro

Here is a quick look at how the core local background agent (app.py) listens to filesystem events and triggers local AI processing:

Python
class SlothBudgetAgent(FileSystemEventHandler):
def on_created(self, event):
if event.is_directory or not event.src_path.endswith((".txt", ".csv")):
return

    time.sleep(1.0)
    file_path = event.src_path

    # Extract features via local Gemma 3 1B & classify via TabPFN-3.5
    with open(file_path, "r", encoding="utf-8") as f:
        raw_text = f.read()

    # Agent processes and files away automatically!

How I Built It
SlothBudget Pro is built on a high-performance hybrid edge-and-cloud architecture optimized to run smoothly even on consumer hardware (like an 8GB RAM Windows PC):

Unstructured Text Extraction (Local Edge): Google's open-weight Gemma 3 (1B) model runs locally via Ollama CPU, extracting structured transaction JSON (merchant, amount, context) from raw text.

Tabular Machine Learning (Local Edge): Prior Labs' TabPFN-3.5 tabular foundation model classifies spending patterns (Food, Bills, Shopping, Travel) and detects anomalies on the extracted features.

Observability & Tracing: Integrated Sentry Agent Tracing measures multi-model execution latency, token throughput, and agent error spans.

Cloud Frontend (Render): The interactive Streamlit dashboard is hosted globally as a live web service on Render, allowing remote viewing and file uploads.

Why Does Open Innovation Matter?
Open innovation made it possible to build a financial assistant with absolute cryptographic privacy. Financial records and bank SMS logs are deeply sensitive; using a closed, cloud-based API would mean sending personal spending data to third-party servers.

By leveraging open-weight models like Gemma 3 and local foundation models like TabPFN-3.5 running entirely offline on consumer hardware, user data stays 100% private, costs $0 to operate, and requires zero internet dependency.

My Agent Session
Sentry Agent Tracing Dashboard: Monitored local agent execution times, ensuring gemma3:1b inference and TabPFN classification completed in under 2 seconds per file.

The Lazy Payoff
This week, I dumped a messy folder of 15 raw transaction text logs into my Unsorted directory. Instead of spending 20 minutes copying amounts into a spreadsheet, SlothBudget Pro quietly processed every file in the background, fired a Windows notification for each categorization, and updated my Render dashboard instantly. Total manual effort: 0 seconds.

Prize Categories
🏆 Slow and Steady (Main Challenge Winner)

☁️ Best Use of Render (Live web frontend deployment)

📊 Best Use of TabPFN (Tabular classification using TabPFN-3.5)

🧠 Best Use of Gemma (Local feature extraction via Gemma 3 1B)

🔍 Best Use of Sentry Agent Tracing (Agent span and latency telemetry)

🌐 Best Use of a .tech Domain (Custom domain routing)