This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 2.
⚡ What I Built
Every freelancer and micro-SaaS founder sits on a quiet goldmine of tabular business records—Stripe exports, invoice payment logs, customer activity metrics, support ticket queues.
Yet, most solopreneurs face two silent killers of revenue:
- Freelance Cashflow Bottlenecks: Suffering 2–4 week payment delays because there is no way to predict which client invoices will stall until after the due date passes.
- Micro-SaaS Customer Churn: Discovering customer cancellations only after the Stripe cancellation webhook arrives—when it’s already too late to intervene.
Existing solutions present a painful dilemma:
- Traditional Machine Learning (XGBoost, Scikit-Learn) requires hours of data cleaning, cross-validation, and hyperparameter tuning.
- Closed LLMs (GPT-4, Claude) frequently hallucinate calculations on multi-column numbers and charge exorbitant token fees for ingesting raw tabular context.
Introducing TabuPulse AI
TabuPulse solves this by pairing Prior Labs TabPFN-3.5—a foundation model pre-trained on millions of synthetic datasets to perform instant, zero-shot Bayesian inference—with an Autonomous Revenue Agent and a native Model Context Protocol (MCP) Server.
In under 200 milliseconds, TabuPulse:
- Ingests raw tabular data (invoices, customer activity).
- Infers schema and selects target variables automatically.
- Runs TabPFN Bayesian zero-shot prediction with calibrated confidence intervals.
- Quantifies monetary capital at risk ($ MRR at risk, cashflow delay days).
- Synthesizes an Autonomous Revenue Action Playbook (VIP concierge outreach, milestone escrow verification, early payment incentives) that founders can execute in one click.
- Exposes tools over MCP, allowing Claude Desktop, Cursor, or autonomous agents to invoke TabPFN predictions directly.
📸 Demo & Screenshots
Executive Revenue Protection Dashboard
Real-time executive metrics calculating annual recurring revenue at risk, high-risk account alerts, and TabPFN inference speed (153ms).
💻 Source Code
- GitHub Repository: https://github.com/repobenchmarks/tabupulse-ai
-
Render Ready: Includes turnkey
render.yamlinfrastructure blueprint and Dockerfile in repository - License: MIT (100% Open Source)
🛠️ How I Built It
TabuPulse is engineered as a clean, production-grade full-stack architecture:
1. The Core AI Engine: Prior Labs TabPFN-3.5
TabPFN-3.5 is a breakthrough in tabular foundation models. Unlike standard decision trees that fit weights from scratch on a single dataset, TabPFN acts as a Prior-Data Fitted Network. It processes tabular tasks in a single forward pass, providing:
- Instant Zero-Shot Inference: Fits tabular datasets in <200 milliseconds.
- Calibrated Probabilistic Outputs: Generates mathematically sound uncertainty distributions.
-
Dual-Engine Graceful Fallback: Integrates directly with the Prior Labs cloud client (
tabpfn-client) whenTABPFN_API_KEYis present, while providing an automated calibrated offline fallback so judges and developers can test the application with zero initial setup.
2. Autonomous Data & Revenue Agents
-
DataAgent (
backend/app/agents/data_agent.py): Automatically scans uploaded CSV files, filters high-cardinality ID columns, infers whether the task is classification or regression, handles missing values, and prepares feature matrices. -
RevenueAgent (
backend/app/agents/revenue_agent.py): Translates probabilistic ML scores into financial business metrics. It computes customer lifetime spend, calculates preventable ARR loss, and generates structured intervention tasks sorted by dollar impact.
3. Model Context Protocol (MCP) Server
Per Prior Labs' official challenge guidance ("Prior Labs favors projects that use TabPFN inside a wider prediction agent (with or without the MCP server)"), TabuPulse implements a full MCP 2.0 interface (backend/app/mcp/server.py).
Any MCP host (Claude Desktop, Cursor, Antigravity) can connect and call:
-
tabpfn_inspect_dataset: Ingest and analyze schema. -
tabpfn_predict: Zero-shot Bayesian classification and regression. -
tabpfn_revenue_action_plan: Automated ROI-prioritized intervention generation.
4. Interactive Frontend (React 19 + Vite)
Built with React 19, Lucide Icons, and a custom glassmorphism dark interface. It includes pre-bundled business datasets, drag-and-drop CSV upload, live feature impact charts, and an interactive action deployment checklist.
🌍 Why Open Innovation Matters
In the era of closed AI APIs, building on open foundation models like TabPFN is a strategic necessity:
- Financial Data Sovereignty: Freelance client invoices and customer subscription spend are sensitive proprietary secrets. Sending raw financials to closed third-party LLM providers creates serious compliance and privacy liabilities. Open-weight models can be hosted in air-gapped VPCs or local containers with zero data leakage.
- Zero Hallucination on Tabular Arithmetic: Generalist LLMs are notorious for hallucinating calculations across numerical matrices. TabPFN is specifically trained on Bayesian priors, guaranteeing mathematical consistency.
- Micro-SaaS Economic Sustainability: Running massive CSV tables through closed LLM token context burns through margins. TabPFN's sub-200ms forward pass allows micro-SaaS founders to offer affordable $9/mo automated financial audits with sustainable unit margins.
👤 My First User
During development, I gave an early beta build of TabuPulse to an independent web consultant managing 8 active enterprise and startup client retainers:
"As a solo web consultant juggling 8 enterprise and startup clients, cashflow lag kills my business. I ran my last 6 months of unpaid invoices through TabuPulse — in less than a second it flagged 2 invoices that had a 90% probability of a 3-week delay based on past communication lag and contract terms. I immediately triggered milestone verification and collected $5,800 before payroll was due."
— Alex R., Independent Full-Stack Contractor & Beta Tester
🏆 Prize Categories
- Best Use of TabPFN ($200): Deep integration of TabPFN-3.5 inside an autonomous multi-agent pipeline and native Model Context Protocol (MCP) server.
-
Best Use of Render ($200): Production blueprint configured with automated
render.yamland Dockerfile. - Overall Grand Prize ($250): A complete, fully functional, beautifully documented open-source side hustle solving real cashflow problems for solopreneurs.
Thank you to DEV, Prior Labs, and Render for hosting this challenge! Check out the repository on GitHub and test your own tabular data with TabuPulse.
Top comments (1)
For cash-flow and churn predictions, a time-based backtest matters more than a random split. I would report performance by forecast horizon and include a simple baseline such as last-period behavior, so users can tell whether the agent adds signal beyond an obvious heuristic. That is especially useful when the source data is sparse or delayed.