DEV Community: Sushan Shetty

🌄FieldCast AI — Before You Go, Know What the Outdoors Holds

Sushan Shetty — Sat, 10 Oct 2026 11:39:54 +0000

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

🌄 Know Before You Go. Experience More Outside

🌲The outdoors is unpredictable.

A hiker may underestimate their return time. A photographer may drive to a viewpoint only to find it covered in fog. A runner may start at a pace they cannot sustain. A camper may miscalculate how much fuel they need.
Existing apps can provide maps, weather forecasts, activity tracking, and general recommendations. But making a good decision often requires bringing that information together with the particular activity, environmental conditions, and a person's own history.
FieldCast AI explores a different approach: predicting what might happen before you go outside, so you can make a better decision before it is too late.

🌍What I Built

The platform brings ten specialised outdoor intelligence concepts into one application:

🥾 TrailGuard: Estimate hiking route difficulties and potential route problems.
🌅 SunCast: Find promising sunset viewpoints and viewing times.
🏃 RunPredictor: Estimate whether a planned running pace is sustainable.
⏱️ TrailReturn: Estimate hiking return times using route progress and daylight.
🏔️ ViewCast: Assess whether a scenic destination is likely to offer a worthwhile view.
🚲 RideCast: Compare cycling times and expected comfort conditions.
🐦 BirdForecast: Estimate which bird species may be encountered.
🌱 GardenEye: Help anticipate plant-care needs.
🏕️ CampFuel: Estimate camping fuel requirements.
🌦️ PlanShield: Assess the feasibility of an outdoor plan and compare alternatives.

Rather than creating ten separate applications, the goal is to build one platform where the appropriate intelligence module is selected according to the activity.

A simple example

🏔️Imagine planning a morning mountain hike.

Instead of looking at separate websites and making every estimate manually, the intended workflow brings together route distance, elevation, daylight, weather, personal hiking pace, and scenic visibility.

The purpose is not to guarantee what will happen. It is to help people make better-informed outdoor decisions.

🔗 Demo

🚀 Open FieldCast AI
Live Demo

The frontend is deployed on Render.

💻 Code

View the FieldCast AI repository on GitHub

GitHub Repository: View FieldCast AI

The repository contains the React + Vite application source code and project configuration.

🛠️ How I Built It

I started with a React + Vite web application using JavaScript and CSS. I built it with Google Antigravity as my AI coding environment, then published the frontend through GitHub and Render. For the open-weight AI side, I used Tinker to work with open models instead of relying on a single closed provider.

Technology	Role
React	Component-based application interface
Vite	Development server and production build
JavaScript	Application interactions and logic
CSS	Responsive layouts and visual design
Google Antigravity	AI-assisted coding environment used to build and iterate on the app
Tinker	Working with and adapting open-weight models for the prediction layer
Git and GitHub	Version control and source repository
Render Static Site	Public frontend hosting

I designed FieldCast AI as a shared platform rather than ten unrelated applications. The intention is to reuse activity history, environmental context, prediction explanations, and outcome recording across the relevant modules.

🧠 The AI architecture

The intended AI layer uses suitable open-weight models and prediction techniques according to the task.

Numerical forecasting may need a different model from natural-language interpretation or image analysis. The project is designed to allow the appropriate model to be selected and replaced rather than forcing every task through one large language model.

Implementation status: The public frontend is deployed. The open-weight AI integration and validated prediction engine must be demonstrated separately before this submission can claim that these capabilities are working.

🌱Why Does Open Innovation Matter?

Outdoor activities can involve sensitive personal information: location history, fitness records, photographs, travel patterns, and daily routines.

Open-weight models offer the flexibility to explore different approaches without committing the entire application to a single closed AI provider.

For FieldCast AI, the intended benefits are:

🔄 Model choice: Experiment with and replace suitable open-weight models.
🔒 Privacy: Keep personal history on the user's device wherever practical.
📡 Offline potential: Explore local inference for locations with poor connectivity.
🔍 Transparency: Explain the factors contributing to a prediction.
🎛️ Control: Adapt the prediction pipeline instead of depending entirely on a proprietary AI service.

These are design goals that need to be demonstrated through the actual implementation. A hosted model does not automatically provide offline operation, and personal data sent to a hosted backend still reaches that server.

The principle behind the project is to use AI as a decision-support tool while keeping the architecture understandable, replaceable, and open to experimentation.

🏆 Prize Categories

Best Use of Render: FieldCast AI's frontend is built with React and Vite and deployed as a Render Static Site. Render hosts the public demo at fieldcast-ai.onrender.com, which took the project from a GitHub push to a live app without any server management.

Best Use of Tinker: I plan to use Thinking Machines' Tinker to fine-tune an open-weight model for outdoor planning and context-aware recommendations in FieldCast AI. I aim to compare it with the base model to measure improvements in prediction quality, response latency, and cost.

Final Reflection

💭 Final Thoughts

Building FieldCast AI taught me that the hard part isn't showing outdoor data. It's helping someone decide quickly, then get out of their way. I'm still early, and most of the ten modules need real data, testing, and open-weight AI integration before their predictions can be trusted.

Plan thoughtfully. Understand the uncertainty. Step outside. 🌿

🙏 Thank you for reading. If you try the demo, I'd love to hear your thoughts and ideas on how FieldCast AI could make outdoor adventures better.💖

📒LedgerFriend — From Everyday Transactions to Smarter Books.

Sushan Shetty — Sat, 03 Oct 2026 11:56:52 +0000

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

Describe the transaction. Review the entry. Understand your books.

LedgerFriend is an experimental bookkeeping assistant for Indian sole proprietors, built for the Hacktoberfest 2026 Weekend Challenge: Build for a Friend.

The idea is simple: a small-business owner should be able to describe a transaction in ordinary language, review the proposed accounting treatment, and generate organised books without manually repeating the same information across multiple reports.

AI suggests the transaction details. Fixed accounting rules generate the entries. A human approves what gets posted.

⚠️ Prototype only. Not production accounting software, tax advice, or a replacement for a Chartered Accountant.

👤 Who I’m Building For

LedgerFriend is designed for an Indian sole proprietor who finds everyday bookkeeping difficult to organise and wants to understand how ordinary business transactions become accounting records.

I have kept the person's identity and business details private rather than inventing a testimonial or publishing personal financial information.

The intended workflow is based on a common problem:

Recording transactions in everyday language is easier than translating them into journals, ledgers, and financial statements.

A user handover and detailed feedback session is still pending.

🔗 Demo

🚀 Open the LedgerFriend Demo

Please use fictional transactions only.

💻 GitHub

View LedgerFriend on GitHub

The repository contains the project source code and setup information.

👤 Who I’m Building For

The intended user is an Indian sole proprietor who is a Friend, Who finds it difficult to keep everyday transactions organised and translate them into accounting records.

I do not want to invent a user testimonial. The project becomes more useful when its design is based on one real person’s workflow.

🛠️ How I Built It

I started with a lightweight, single-page application using:

Component	Role
HTML	Application structure and forms
CSS	Responsive dashboard and report styling
JavaScript	Validation, posting rules, reports, and AI requests
Browser localStorage	Prototype data persistence
Ollama	Local model serving
Qwen2.5	Open-weight language model for transaction parsing
VS Code	Editing and local development
Python HTTP server / VS Code Live Server	Serving the interface locally

I installed and experimented with:

qwen2.5:3b
qwen2.5:1.5b

The smaller model was explored to reduce hardware demands. Smaller size does not guarantee adequate accounting classification accuracy.

AI-assisted development tools

I used AI assistance to explore the idea, generate and revise code, and troubleshoot integration problems.

Tools used during the development process included:

Backboard: I used Backboard to create the initial project structure and HTML foundation for LedgerFriend. It helped turn the initial concept into the application's page structure, forms, and overall interface layout, which I then developed and refined further.
Antigravity: I used Antigravity during development to help with implementation and debugging, especially while working through application behaviour and integration issues. It helped me iterate faster while testing different approaches.
ChatGPT and Claude: ideation, implementation guidance, and debugging assistance.
Thinking Machines: brainstorming, refining the bookkeeping workflow, exploring edge cases, and improving implementation ideas. Development assistance is distinct from the application’s runtime dependencies.

The documented local inference path is Ollama + Qwen. I should only claim a Backboard runtime integration or a partner-category feature if it is present in the submitted implementation.

Keeping AI away from authoritative posting

The intended flow is:

Description → AI suggestion → Validation → Human review → Rule-based entry → Reports

The model proposes a supported transaction type and extracts fields such as amount, date, and reference.

Application code then selects accounts from fixed rules and calculates with integer paise. The owner must review and confirm before posting.

★This does not eliminate mistakes. An incorrectly classified transaction can still balance.

Equal debits and credits are a necessary check—not proof that the accounts are correct.

Unclear notes belong in a review queue. That queue is not a suspense account, and it does not affect the books.

✨ Prototype Features

📝 Transaction Entry

Plain-language transaction descriptions
Local AI integration through Ollama
Manual entry when AI is unavailable
Fixed list of supported transaction types
Journal preview and explicit confirmation before posting

📊 Accounting Views

Journal
General Ledger
Trial Balance
Draft Trading Account
Draft Profit & Loss Account
Draft Balance Sheet

⚠️ Scope and Limitations

This prototype is intentionally narrow:

INR only
Sole proprietorships only
First bookkeeping period with zero opening balances
Periodic inventory accounting
No GST/TDS calculations or statutory filing

🏆 Prize Categories

Best Use of Backboard Build with R-CLI, Backboard's open-source terminal coding agent, compare open-weight models through a single Backboard API key, or give an open-source project's assistant memory and RAG.
Best Use of Tinker Used Tinker to experiment with model behaviour and improve LedgerFriend's AI-assisted bookkeeping workflow. Explored model outputs, edge cases, and ways to make transaction interpretation more reliable.

Final Reflection

I started with a simple question:

“Can AI create the accounts?”

Building LedgerFriend led me to a better one:

“How can AI make bookkeeping easier without silently becoming the authority?”

That question became the foundation of LedgerFriend.

And that is the project I want to keep building.

❤️ Thank You

Thank you for taking the time to explore LedgerFriend!

This project was a learning experience in building a bookkeeping application, integrating local AI, debugging real-world problems, and separating AI suggestions from accounting rules

🌍 Deep-Time Mirror: An AI Lens into Our Ecological Soul For Earth Day Edition.

Sushan Shetty — Fri, 17 Apr 2026 16:59:49 +0000

This is a submission for Weekend Challenge: Earth Day Edition (April 22)

What I Built

I built the Deep-Time Mirror, an immersive, AI-powered "temporal lens" that allows users to witness the environmental evolution of any location on Earth.

The goal was to move beyond cold, flat data and create an emotional connection with the land. The Mirror takes you through three distinct epochs:

The Deep Past (~10,000 BCE): A cinematic look at the primeval ecosystem before modern intervention.
The Present (The Anthropocene): A gritty, real-time look at current urban and climate impacts, including a "Recent Pulse" with local environmental events from the last 7 days.
The Restored Future (~2100 CE): A hopeful, Solar-punk vision where humanity has successfully restored biodiversity and integrated technology with nature.

🌱 The Earth Auditor (Gaia Corp)

Beyond observation, I built a personal accountability system:

Users “confess” their daily habits to a brutally honest AI Earth.
The system generates a Planetary Stability Score (0–100).
It breaks down impact across:
Energy ⚡
Transport 🚗
Waste ♻️
Water 💧
Diet 🍽️

All of this is visualized through an interactive radar chart, turning abstract habits into tangible environmental impact.

Demo

Experience the Mirror live here:

👉 Deep-Time Mirror Live Demo

Code

The full source code and documentation can be found here:

sushanshetty2601-pixel / deep-time-mirror

🌍 An AI-powered 3D cinematic portal into Earth's ecological evolution. Visualize the primeval past, the modern Anthropocene, and a solar-punk future for any location. Built with Gemini AI, React, and Three.js for the Earth Day 2026 Challenge.

🌍 Deep-Time Mirror

Piercing the veil of time for Earth Day 2026.

Deep-Time Mirror is an immersive, AI-powered ecological portal that visualizes the environmental history and restored future of any location on Earth. Built for the Earth Day Challenge, it combines high-performance 3D graphics with real-time generative storytelling to bridge the gap between our primeval past and a sustainable solar-punk tomorrow.

⚡ The Experience

🔭 The Deep-Time Mirror

Temporal Synthesis: The AI stabilization engine kicks in, generating a cinematic, full-screen 3D visual of your location.
Deep Past (~10,000 BCE): Witness the land in its untouched, primeval glory. Meet ancient species and see the world before the Anthropocene.
The Present: A gritty, realistic look at current ecological challenges, complete with a "Recent Pulse" that grabs local environmental data from 1 day and 1 week ago.
Restored Future (~2100 CE): A solar-punk vision of restoration. Witness how architecture and nature can coexist…

View on GitHub

How I Built It

My technical approach focused on Temporal Synthesis—using AI to ground creative visuals in scientific fact.

Google Gemini 1.5 Flash: I used Gemini to act as a "Temporal Oracle." It performs real-time searches to gather ecological data about the user's location, then generates structured JSON data for the narratives and climate statistics.

3D Cinematic Visuals: I utilized a generative visual pipeline that creates a full-screen, immersive background for every location. This "Art-First" approach ensures that users see the impact of climate change before they read about it.

React 19 & Three.js: The core interface uses React 19 for high-speed state management and Three.js (via React-Three-Fiber) to create the "Temporal Sphere" at the heart of the experience.

Design Philosophy: I used a high-end editorial/luxury aesthetic with Tailwind CSS v4, utilizing Glassmorphism and specialized scanline filters to give the app a "Retro-Futuristic Science" feel.

Data Visualization: I integrated Recharts to create custom-themed radar charts and built SVG Progress Dials to visualize the Planetary Score. These are fully responsive, data-driven components that transform AI output into actionable insights.

Advanced State Management: Leveraging React 19 hooks and Framer Motion, I created seamless “Temporal Transitions” that make switching between the Mirror and the Auditor feel fluid and cinematic.

Prize Categories

Best Use of Google Gemini

🌍 Grounded Real-Time Search: To find hyper-local environmental changes from the last 24 hours and 7 days.
🧠 Complex Reasoning: To generate scientifically plausible “Restored Future” scenarios based on geography and climate constraints.
🎭 Brutally Honest Personality Logic: Gemini is engineered as a strict planetary auditor, not a passive chatbot. It detects harmful habits (like plastic use or waste) and dynamically assigns an Earth Grade (A+ to F), creating a strong moral feedback loop.
📈 Multi-Dimensional Data Synthesis: It outputs structured data that powers the Planetary Stability Score and a 5-axis environmental radar chart, ensuring every user interaction produces unique, data-driven insights.

✨ Special Features

🌐 High-Fidelity Details

Custom Temporal Cursor: A glowing Earth orb with an orbit ring and subtle leaf trail that reacts dynamically to user interactions.
Live Pulse Integration: The landing experience includes simulated global activity (12k+ locations, 3.2M trees restored), making users feel part of a global movement.
Glassmorphism Design System: Every UI layer uses blur, transparency, and subtle borders to mimic a futuristic “Gaia Corp” terminal.
Cinematic Transitions: Moving between the Mirror and Auditor feels like navigating a high-end sci-fi interface.

🍃 A Message for Earth Day 2026 (April 22)

Earth Day is a reminder that we are not just residents of this planet—we are its custodians.

The Facts We Face:

Carbon Consciousness: CO₂ levels are currently at their highest in over 4 million years. Our “Deep Past” view reminds us of the baseline we are trying to return to.
Restoration is Possible: Scientists estimate that restoring just 15% of converted lands in priority areas could avoid 60% of expected extinctions. My “Future” view is a digital blueprint of this possibility.
Every Day Matters: The “Recent Pulse” in the app monitors changes from just 24 hours ago. Climate action isn't a long-term goal—it’s a daily rhythm.

Built with 💚 for the planet by: Sushan Shetty (sushanshetty2601-pixel)
Thank you 😇

✨ What I Built for the DEV April Fools Challenge: 🎵 Shake-to-Volume Music Player + 🐍 Shake-Controlled Snake Game

Sushan Shetty — Wed, 08 Apr 2026 17:20:03 +0000

🎵🐍 Shake-to-Volume Music Player + Tilt-Controlled Snake: Double Chaos Edition

A hilariously useless double feature where EVERYTHING is controlled by shaking/tilting your device!

What I Built

I created the world's most frustrating double feature:

A music player where volume ONLY works by shaking your device
A Snake game controlled entirely by tilting your phone or violently moving your mouse

Both features are completely useless, intentionally difficult, and will make you question reality.

Demo

🔗 Live Demo: https://silly-dev.preview.emergentagent.com/
🔗 GitHub: https://github.com/sushanshetty2601-pixel/shake-chaos-player

Screenshots

Features (All Intentionally Useless)

🎵 Shake-to-Volume Music Player

Shake Detection: Uses accelerometer on mobile, requires violent mouse movements on desktop
Volume Control: NO buttons, NO sliders - ONLY shaking works
Volume Decay: Stop shaking? Volume drops to zero. Keep shaking forever!
Shake Meter: Real-time visualization of your pathetic shaking intensity
Stats Tracking: Counts every single shake and scores your desperation
🤖 AI Shake Coach: Google AI roasts your shaking technique in real-time
Calibration Mode: 5-second test that judges if your shaking is "worthy" (spoiler: it's not)
Shake History: Database permanently records your shameful shaking attempts

🐍 Shake-Controlled Snake Game

Tilt Controls: On mobile, tilt device to change direction. On desktop, move mouse away from screen center.
No Arrow Keys: Traditional controls are for cowards
Game Over Messages: Sarcastic AI roasts like "😴 Pathetic tilting skills"
Score Tracking: Watch your failure in real-time
Classic Snake: Eat food, grow longer, inevitably crash into yourself while trying to tilt properly

Why This Is Hilariously Useless

Music Player Problems:

Want to turn up the volume? Shake violently for 30 seconds straight.
Listening on a bus? Everyone thinks you're having a seizure.
Desktop users must thrash their mouse around like a possessed etch-a-sketch.
Stop shaking for 2 seconds? Silent. Forever shaking or forever silent.

Snake Game Problems:

Playing Snake is already hard. Playing it by TILTING is impossible.
Desktop users must remember which quadrant their mouse is in.
Every game ends in 10 seconds with a sarcastic AI insult.
You'll look ridiculous tilting your phone back and forth.

Combined Chaos:

Try playing Snake while the music volume keeps resetting because you stopped shaking.
Ultimate multitasking nightmare.

How I Used Google AI (Gemini 3 Flash)

I integrated Google Gemini 3 Flash API to create an "AI Shake Coach"

Implementation:

Backend: FastAPI endpoint /api/shake-coach
Model: gemini-3-flash-preview via Google Generative AI API
Integration: Used Emergent's LLM integration library for seamless Gemini access

What the AI Does:

Analyzes Shake Data: Takes intensity, count, score, and device type
Generates Sarcastic Feedback: Creates custom roasts based on performance
Assigns Roast Levels: Pathetic → Mediocre → Acceptable → Legendary
Provides "Useless Tips": Like "Shake like you're trying to wake up your phone's ancestors"

Sample AI Responses:

"Congratulations on vibrating your desk with the raw power of a drowsy hamster..."
"At this rate, you'll reach 'audible' volume by the next solar eclipse"
"I've seen more intensity from a leaf of lettuce in a mild breeze"

Why This AI Usage Is Perfect for the Challenge:

The AI takes something already useless (shaking for volume) and makes it MORE useless by roasting you
Gemini's creative language generation makes every roast unique
It's meta-useless: AI judging your performance in a completely pointless task

Tech Stack

Frontend: React, Tailwind CSS, Web Audio API, DeviceMotion API
Backend: FastAPI, Python
Database: MongoDB (to forever store your shame)
AI: Google Gemini 3 Flash via Generative AI API
Deployment: Emergent platform
Motion Detection: DeviceOrientationEvent (mobile), MouseEvent (desktop)

How to Try It

Mobile:

Visit the live demo on your phone
Click "Play Music"
Shake your phone aggressively
Watch volume go up (and immediately drop when you stop)
Try the Snake game by tilting
Get roasted by AI

Desktop:

Visit on your computer
Move mouse violently in all directions
Or spam keyboard keys
Try controlling Snake by moving mouse to screen edges
Fail spectacularly

Code Highlights

Shake Detection (Mobile):

window.addEventListener('devicemotion', (event) => {
  const { x, y, z } = event.accelerationIncludingGravity;
  const acceleration = Math.sqrt(x * x + y * y + z * z);
  if (acceleration > 20) {
    const intensity = Math.min(acceleration / 50, 1);
    setVolume(Math.floor(intensity * 100));
  }
});

AI Shake Coach (Backend):

chat = LlmChat(
    api_key=os.environ.get('EMERGENT_LLM_KEY'),
    session_id=f"shake-coach-{uuid.uuid4()}",
    system_message="You are a sarcastic Shake Coach..."
).with_model("gemini", "gemini-3-flash-preview")

ai_response = await chat.send_message(user_message)

What I Learned

DeviceMotionEvent API is surprisingly powerful (and hilarious)
Gemini 3 Flash is excellent at generating creative sarcasm
People will question your sanity when they see you violently shaking your phone
Making something intentionally useless is harder than making something useful
AI makes everything better, even roasting your terrible shaking technique

Conclusion

This project perfectly embodies the April Fools spirit: it's technically impressive, completely useless, and will frustrate anyone who tries it. The AI integration adds an extra layer of absurdity by judging and roasting your attempts to use an already ridiculous interface.

Try it, shake it, tilt it, and get roasted! 🎉

Built with frustration, bad ideas, and Google AI by @sushanshetty2601-pixel

This app is intentionally useless. Created for DEV April Fools Challenge 2026.