DEV Community: Parth

Retention AI

Parth — Mon, 19 Oct 2026 07:29:10 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

RetentionAI is an executive-grade subscription churn intelligence tool that turns passive cancellation risk into automated, revenue-protecting business actions.

The Problem: Subscription and micro-SaaS founders lose thousands of dollars each month to silent churn. By the time a customer clicks "Cancel Subscription," it’s already too late. Most churn-prediction solutions either require massive datasets and complex MLOps pipelines (XGBoost, hyperparameter tuning loops) or cost thousands per month on enterprise platforms like ChurnZero or Gainsight.
The Solution: RetentionAI leverages TabPFN (PriorLabs' foundational transformer for tabular data) to deliver instant, in-context churn predictions on small-to-medium customer cohorts without any hyperparameter tuning. It flags high-risk accounts, calculates total monthly recurring revenue (MRR) at risk, prescribes targeted retention plays (e.g., automated 20% win-back incentives, fast-track engineering escalation, VIP 1-on-1 reviews), and features an interactive What-If ROI Simulator to forecast exact dollar returns on retention campaigns.
Who would pay for it? Indie hackers, bootstrapped SaaS founders, Customer Success (CS) leads, and RevOps managers running subscription products on Stripe, Paddle, or Lemon Squeezy who need proactive churn prevention without hiring an ML engineer.

Demo

💻 Local Preview: Run streamlit run app.py at http://localhost:8501

Code

The complete codebase is open-source and pre-configured for 1-click cloud deployment:

parthbadgire-code / retentionai

RetentionAI: Subscription Churn Predictor using TabPFN tabular foundation model

📉 RetentionAI: Subscription Churn Predictor

An enterprise-grade, Render-ready Streamlit web application powered by TabPFN (PriorLabs' foundational transformer model for tabular datasets) to predict subscriber churn risk and automate retention actions.

🌟 Key Features

TabPFN Client Integration: Direct connection to PriorLabs' in-context tabular foundation model via tabpfn-client.
Zero-Config Demo Mode: Ships with 200 realistic synthetic subscription customer records and a default 25-account evaluation cohort so it functions right out of the box.
Custom CSV Ingestion: Upload any subscriber dataset with support for custom customer IDs and dynamic column mapping.
Dynamic Executive KPIs:
- 👥 Customers Analyzed
- 🚨 High-Risk Accounts (with dynamic risk threshold)
- 💰 Monthly Revenue at Risk ($)
Automated Retention Playbook: Prescribes targeted business actions (e.g. 20% win-back discounts for inactivity >21 days, urgent CS escalation for support tickets ≥3, VIP executive calls for high-spend tiers).
Interactive Visualizations: Churn distribution histograms, inactivity vs. churn scatter charts, and revenue breakdown by…

View on GitHub

Repository link: https://github.com/parthbadgire-code/retentionai

Key components:

app.py: Full Streamlit web application featuring TabPFN integration, Scandinavian pastel design system, and the What-If ROI Simulator.
generate_data.py: Synthetic customer cohort generator with realistic feature correlations (seed=42).
render.yaml & Procfile: Render Blueprint specifications for zero-config cloud hosting.
.streamlit/config.toml: Custom light porcelain theme and minimal toolbar settings.

How I Built It

RetentionAI is built around the modern Python open-source AI ecosystem:

Tabular Foundation Model (TabPFN): Built using tabpfn-client. Unlike classical gradient-boosted trees that require training loops and hyperparameter search, TabPFN operates as an in-context transformer trained on millions of synthetic datasets. It produces calibrated Bayesian posterior probabilities directly in a single forward pass.
Streamlit (Scandinavian Pastel UI): Built with streamlit using a bespoke executive dashboard design system inspired by tools like Linear and Stripe. It avoids dark "neon-AI" tropes in favor of soft porcelain surfaces (#F8FAFC), crisp white cards, muted sage/terracotta indicators, and Plus Jakarta Sans typography.
Synthetic Dataset Engine: Built using pandas and numpy to generate realistic customer records with real-world churn drivers:
- Days elapsed since last login ($1$ to $45$ days)
- Tier spend ($15, $30, $99 MRR)
- Support complaints (Poisson distribution, $\lambda = 1.5$)
- Engagement session duration (Normal distribution, clipped $1$ to $60$ mins)
Interactive What-If Retention ROI Simulator: A dynamic scenario modeling engine that lets teams adjust conversion rates (discount win-back %, support resolution success %, VIP review save %) to model projected preserved MRR, annualized ARR protected, campaign costs, and net ROI multiples in real time.
Calibrated Fallback Engine: Built with scikit-learn (HistGradientBoostingClassifier) to enable seamless offline and demo testing even without an active API key.

Why Does Open Innovation Matter?

Over 80% of critical enterprise data lives in tables, spreadsheets, and relational databases. Yet the AI revolution has disproportionately focused on closed LLM APIs, leaving tabular data reliant on fragmented, complex pipelines or vendor lock-in.

Open foundation models like TabPFN prove that foundation model architectures can solve structured business problems with zero configuration. Open innovation makes enterprise-grade predictive intelligence accessible to solo creators and bootstrapped businesses, allowing them to retain customers and compete with venture-funded giants on equal footing.

My Agent Session

This project was designed, implemented, refined, and deployed through an end-to-end pair-programming session with Antigravity (Google DeepMind's advanced coding agent):

Generating realistic synthetic data cohorts with non-deterministic logistic risk distributions.
Integrating the PriorLabs TabPFN client authentication and fallback mechanisms.
Iteratively redesigning the interface from dark neon styling to a clean, muted pastel executive theme.
Architecting the interactive What-If Retention ROI Simulator and Render Blueprint files.

My First User

I tested RetentionAI with a fellow indie SaaS developer managing ~120 active subscriptions. By running their recent activity export through RetentionAI:

The Result: The model immediately flagged 6 accounts with high support ticket counts and declining session lengths that hadn't yet cancelled.
The Action: The founder scheduled proactive check-ins and offered feature walkthroughs, preventing an estimated $270/mo ($3,240 ARR) in silent churn.
Their Feedback: "Normally I only notice someone churned when Stripe sends a cancellation webhook. Seeing the exact dollar value at risk before they leave turns churn prevention into a routine Monday morning task."

Prize Categories

Open-Source AI Challenge: Side Hustle (Best revenue-generating tool built with open AI)
Best Use of Tabular Open-Source AI / TabPFN

TouchQuest: The Anti Doomscroll AI Companion

Parth — Mon, 12 Oct 2026 06:16:49 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Modern apps and algorithmic feeds are engineered for infinite retention: notifications, autoplay loops, and endless dopamine scrolls keep us locked inside, staring at glass screens.

TouchQuest flips the paradigm. Instead of keeping you glued to a feed, TouchQuest gives you a structured reason to step outside, pocket your device, and engage your senses with the real living world.

Designed mobile-first with an earthy forest aesthetic (forest greens, warm cream, muted sage), TouchQuest serves outdoor enthusiasts, students, remote workers, and anyone looking to break screen fatigue through micro-adventures:

AI Outdoor Mission Generator: Generates focused, sensory-grounded quests for 15, 30, or 60 minutes based on your local terrain (urban park, forest trail, backyard, or riverbank). Each mission guides your physical attention—listening for wind shear in tree canopies, feeling bark texture, or hunting sunlight patterns—with explicit reminders to keep your phone in your pocket between milestones.
Nature Scanner: A multimodal open-weight vision scanner. Capture or upload a specimen photo to identify wild plants, trees, fungi, and natural organisms with scientific taxonomy and field observation tips.
BirdCall AI: A dedicated bioacoustics classifier interface for recording avian songs and identifying species vocalizations in the field.
Explore Nearby (Privacy-Preserving Cartography): An interactive map built with Leaflet + OpenStreetMap that queries real public green spaces, botanical sanctuaries, and wetlands via the Overpass API. It calculates real-time walking distances and features a 1-tap "Start Quest Here" launcher with zero surveillance advertising or location tracking sold.
Field Nature Journal & Auth: A personal field notebook powered by Supabase Auth & PostgreSQL Row-Level Security (plus a resilient offline Guest mode). It tracks cumulative minutes spent offline, logs sensory reflections, and preserves biodiversity sightings.

Demo

Live Application: https://touchquest.vercel.app
Features to try live:
- Launch a 15-minute sensory quest on the Missions tab.
- Explore nearby parks and nature reserves on the Nearby Parks map using your GPS location.
- Test the Nature Scanner by photographing leaves or trees.
- Log your offline time and reflections in the Field Journal.

Code

The entire codebase is open-source under the MIT License:

parthbadgire-code / touchquest

A mobile-first, AI-powered nature exploration companion for outdoor micro-adventures, biodiversity discovery, and avian bioacoustics.

🌿 TouchQuest • Explore More. Scroll Less.

A mobile-first, AI-powered nature exploration companion designed to break algorithmic dopamine loops, inspire real-world outdoor micro-adventures, identify local biodiversity, and record avian bioacoustics.

Live Application: https://touchquest.vercel.app

🌲 Vision & Problem Statement

Modern feeds and short-form algorithms are tuned for continuous retention. TouchQuest flips the paradigm: instead of keeping you on the glass screen, it gives you a tactile outdoor quest (15, 30, or 60 minutes), guides your attention toward local flora and birdsong, and asks you to pocket the phone.

🛠️ Tech Stack

Framework: Next.js 14+ (App Router) with TypeScript & React 19
Design System: Vanilla Tailwind CSS with custom nature palette (Forest green #123324, Muted sage #88a38d, Warm cream #faf8f2, Earth ochre #c27d38), responsive mobile-first bottom navigation, glassmorphism tokens
Icons: lucide-react
Database & Auth: Supabase (PostgreSQL with Row-Level Security, Auth, Storage)
Maps: Leaflet & OpenStreetMap (Privacy-preserving local…

View on GitHub

Repository: https://github.com/parthbadgire-code/touchquest

How I Built It

TouchQuest is built around a resilient, dual-provider Adapter Pattern that strictly utilizes genuine hosted open-weight models:

Frontend & App Shell: Next.js 16 (App Router), React 19, TypeScript, Tailwind CSS v4, and Lucide icons.
Hosted Open-Weight AI Inference:
- Missions & Vision Scanner: Powered by Groq Cloud running qwen/qwen3.8-27b—a state-of-the-art open-weight multimodal model. Missions use strict JSON schema mode for sub-second generation (~400ms latency), and the Nature Scanner processes base64 image data for botanical and fungal taxonomy.
- Fallback Provider: An adapter fallback to Hugging Face Router v1 (Qwen/Qwen2.5-7B-Instruct and Qwen/Qwen2.5-VL-72B-Instruct).
- Acoustic Serving: Dedicated Cornell BirdNET endpoint hook (HF_AUDIO_ENDPOINT) with 10MB upload guards and graceful standby reporting.
Cartography & Geodata: Leaflet loaded dynamically (ssr: false) with OpenStreetMap tiles. A custom Overpass API engine (src/lib/services/parks-service.ts) queries OSM nodes for leisure=park, nature_reserve, and garden, calculating Haversine distance from the user.
Database & Authentication: Supabase PostgreSQL with Row-Level Security policies safeguarding user profiles, saved quests, sightings, and journal entries, backed by a local storage synchronization layer for offline field resilience.

Why Does Open Innovation Matter?

Nature Exploration Shouldn't Be Gated Behind Proprietary Walled Gardens: When nature apps rely on closed, black-box APIs, users are subjected to price hikes, telemetry lock-in, and sudden shutdowns. Open-weight models ensure that ecological tools remain transparent, accessible, and community-owned.
Accessible Biodiversity Education: Open models like Qwen, BioCLIP, and BirdNET empower developers worldwide to build localized, multilingual taxonomy tools without requiring expensive enterprise credits or credit cards.
Data Privacy in the Outdoors: Location data and personal journaling should remain private. By combining open-weight models with OpenStreetMap and client-side privacy boundaries, TouchQuest proves that AI can enhance our relationship with the natural world without turning nature into an ad-targeting engine.

My Agent Session

This application was engineered through an end-to-end pair-programming session with Antigravity (Google DeepMind), progressing through three structured milestones:

Architectural scaffolding and mobile-first design tokens.
AI adapter auditing, replacing deprecated serverless endpoints with hosted open-weight inference on Groq, and verifying live JSON schema generation.
Leaflet OpenStreetMap integration, Supabase AuthContext implementation, and production deployment on Vercel.

Prize Categories

Touch Grass Challenge (Primary Track): Getting people away from algorithmic dopamine loops and into the physical outdoors through 15/30/60-minute sensory micro-adventures, biodiversity scanning, and outdoor journaling.
Open-Weight AI: Built exclusively with open-weight models (qwen/qwen3.8-27b, Qwen-2.5, BioCLIP, and BirdNET).

Kredence: Decentralized IP Protection for Creators (Hacktoberfest 2026)

Parth — Sun, 04 Oct 2026 13:37:24 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

Kredence is a decentralized, on-chain protocol built on Solana that protects digital creators by establishing immutable proof of originality, providing trustless licensing mechanisms, and maintaining a community-governed dispute resolution system. I built it for a digital artist friend who frequently had their art stolen and reposted without credit. Kredence solves this by using a Commit-Reveal scheme combined with Compressed NFTs (cNFTs) to provide verifiable receipts and a dispute system.

Demo

https://kredence31.vercel.app/

Code

Kredence GitHub Repository

How I Built It

Kredence is built on the Solana blockchain:

Frontend: Next.js 15, React 19, TailwindCSS v4, Web3.js
Smart Contract: Solana, Anchor Framework, Rust
Infrastructure: Metaplex Bubblegum (cNFTs) for state rent optimization, SPL Account Compression, Token-2022, and Pinata (IPFS) for decentralized storage.
In-Browser Hashing: A unique SHA-256 fingerprint of the image is generated locally before it hits the chain.

Why Does Open Innovation Matter?

Open innovation matters because creators need systems they can trust. Centralized platforms can fake timestamps, manipulate data, or simply shut down. By building Kredence as an open-source protocol on a public blockchain, we empower creators to take control of their intellectual property without relying on a walled garden.