DEV Community: Yash Raghubanshi

LedgerGhost: The Air-Gapped AI Bookkeeper That Saves Freelancers $3,000+ at Tax Time

Yash Raghubanshi — Thu, 15 Oct 2026 16:33:57 +0000

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

What I Built

Every freelancer, independent software contractor, and solo agency founder knows the creeping dread of tax season.

Throughout the year, receipts and invoices pile up across email inboxes, camera rolls, PDFs, and client statements. At the same time, freelancers operate under strict Non-Disclosure Agreements (NDAs). You can’t just feed raw client invoices, subcontractor rates, or project memos into commercial cloud LLMs like OpenAI or Anthropic—doing so can breach client confidentiality contracts and leak unreleased project names or private rates to third-party servers.

On top of that, platforms like QuickBooks and FreshBooks charge $35 to $80 every single month, charging even more for automated receipt scanning. For a solo freelancer making $4,000–$10,000/month, that's $500–$1,000 drained on software rent each year. Worse yet, self-employed professionals leave an estimated $3,000+ per year in unclaimed IRS Schedule C tax write-offs because manually categorizing deductions (such as the strict 50% business meals cap, Section 179 hardware write-offs, and 1099 contractor labor) is tedious and confusing.

To solve this, I built LedgerGhost: The 100% Air-Gapped, Zero-Data-Leak AI Bookkeeper & Invoice Auditor for Freelancers.

What It Does

100% Local Receipt Ingestion: Parses messy receipt texts, software bills, travel slips, and contractor invoices locally on your laptop with zero cloud telemetry.
Local Gemma 2 Extraction: Extracts clean structured data (vendor, date, line items, taxes, totals, payment methods) powered by Google's Gemma 2 open-weights model.
Forensic IRS Schedule C Categorization: Automatically classifies expenses into IRS tax lines (Line 18 Office & SaaS, Line 11 Contract Labor, Line 24a Travel, Line 13 Section 179 Hardware) and enforces statutory rules like the IRS 50% deduction limit on business meals.
Ghost Audit™ Anomaly & Leak Scanner:
- Duplicate Billing: Detects duplicate vendor charges across billing cycles.
- Subscription Price Creep: Flags when a SaaS subscription quietly raises prices (e.g., $15/mo up to $21/mo).
- Client NDA & PII Leak Auditor: Scans expense memos for confidential client project codenames or credit card numbers, preventing accidental leaks on client billing.
Client Invoice Generator: Generates professional, clean markdown and plaintext invoices in seconds.
Pure Python Visual Dashboard: Includes an interactive terminal interface built with rich and an air-gapped visual web dashboard built with Python's standard http.server with dynamic SVG charts.

Who Would Pay For It?

Target Customer: Solo software developers, independent consultants, freelance designers, and boutique agencies.
The Business Model: A $49 one-time lifetime license (the beloved Obsidian / TablePlus / Alfred model) or a $9/month local self-hosted tier.
The Value Proposition: For $49, a freelancer saves 10+ hours a month on bookkeeping, stays 100% NDA-compliant, avoids SaaS rent, and reclaims $1,500–$4,000 in legitimate tax write-offs at zero per-token cost.

Demo

Here is LedgerGhost running locally on a laptop terminal:

1. Ingesting Real-World Receipts with Gemma 2


bash
python run.py demo

WildPulse: The Zero-Signal Trail Naturalist That Tells You to Put Your Phone Away

Yash Raghubanshi — Wed, 07 Oct 2026 21:10:35 +0000

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

What I Built

Most modern "nature apps" have a fatal, hypocritical flaw: they demand that you hike out into pristine wilderness, stop in front of a majestic sugar maple or a singing robin, and... stare at a bright glass rectangle for ten minutes.

They make you scroll through text, squint at Wikipedia paragraphs, and look down at your screen instead of looking up at the forest canopy right above your eyes. And the moment you lose cellular reception on a backcountry trail, they crash with an ERR_NETWORK_DISCONNECTED error.

I built WildPulse around one fundamental design law:

The screen must be the shortest part of the outdoor experience.

WildPulse is a zero-signal, offline-first field naturalist and audio companion designed for trail walkers, birders, foragers, and autumn foliage explorers.

The Real-World Experience:

Quick Specimen Snap or Sound Tap: Point your camera at a vibrant autumn leaf, tree bark, or mushroom, or tap to capture 3 seconds of birdsong.
Instant On-Device Identification: An open-source, local-first taxonomic intelligence identifies the specimen in under 1 second without pinging any remote cloud server.
15-Second Earbud Narration: Rather than making you read dense paragraphs on a bright screen, WildPulse speaks a brief, documentary-style audio narration directly into your earbuds.
Immediate Screen Sleep ("Pocket Mode"): The moment narration begins, the screen automatically dims to a restful, minimalist dark display with a gentle pulsing green leaf and a message: > "Audio playing in your earbuds. Put your phone in your pocket and enjoy the canopy."
Touch Grass Scavenger Quests: 5 seasonal nature quests (Autumn Foliage Encounter, Forest Floor Mycota, Keystone Tree Canopy, Songbird Soundscapes, and Trailside Pollinators) turn any walk into an engaging real-world treasure hunt.
Private Local Trail Journal: Observations are recorded to local storage on your device with zero GPS tracking or cloud telemetry, exportable as clean Markdown.

Demo

Live GitHub Repository: https://github.com/Yash990-bit/WildPulse
Offline Mode Tested: 100% operational in Airplane Mode. Try switching off Wi-Fi and mobile data—WildPulse runs completely unhindered with zero dropped frames.

What it Looks Like in the Field:

Field Lens & Quick Presets: Instant botanical taxonomy with confidence metrics.
Audio Field Guide: Hands-free narration delivered with calm naturalist cadence.
Pocket Mode HUD: Auto-dims the display to break screen addiction on the trail.
Autumn Nature Quests: Interactive real-world scavenger hunts that motivate you to hike further.

Code

The entire codebase is open source under the MIT License:

Yash990-bit / WildPulse

🍃 WildPulse: Zero-Signal Offline Field Naturalist & Audio Companion

"Build something with open-source AI at its core that gets people off the screen and into the world."
Submission for Hacktoberfest 2026 DEV Challenge — Week 1: Touch Grass

🌲 The Problem: Why Most "Nature Apps" Fail the Forest

When you hike out to a ridgeline or wander deep into an old-growth forest, two things happen:

Cellular signal drops to zero. The second you hit airplane mode or lose LTE, every cloud-dependent AI assistant (ChatGPT, Gemini Cloud, Claude) crashes with ERR_NAME_NOT_RESOLVED.
Screens ruin the moment. Conventional apps force you to stand still, squint at a bright glass rectangle, read dense paragraphs of text, and look away from the wildlife right in front of you.

WildPulse was built with one cardinal rule:
The screen must be the shortest part of the experience.

✨ What WildPulse Does

WildPulse is a zero-signal, offline-first…

View on GitHub

Core Tech Stack:

Architecture: HTML5 PWA, Vanilla Modern CSS, JavaScript (ES2024).
Audio Synthesis: Native Web Speech API for 100% offline speech, plus ElevenLabs Voice Synthesis API integration for cinematic documentary audio.
Offline Durability: Service Worker (sw.js) with Cache Storage for zero-signal backcountry reliability.
Deployment: render.yaml blueprint ready for 1-click cloud or edge deployment.

How I Built It

To fulfill the "Touch Grass" theme honestly, the application could not rely on traditional cloud API calls that fail when you are miles away from the nearest cell tower.

1. Zero-Signal On-Device Engine

The identification core runs completely inside the client using open-source, local-first classification principles. It parses visual characteristics (such as leaf lobe geometry and autumn anthocyanin/carotenoid pigmentation ratios) and audio frequency bands (isolating high-pitch avian whistles from low-frequency wind rustle).

2. Audio-First Interface with Web Speech & ElevenLabs

To keep eyes off the screen:

Default Offline Engine: Uses the browser's native Web Speech synthesis tuned to a measured, documentary cadence (rate: 0.92, pitch: 0.95). Zero network traffic, zero API keys required, works in pure airplane mode.
Partner Enhancement (ElevenLabs): For users wanting cinematic narration, WildPulse integrates ElevenLabs' Text-to-Speech API with calm naturalist voice profiles to deliver BBC-style field audio guides.

3. PWA Service Worker for Remote Backcountry

WildPulse registers a custom Service Worker (sw.js) that caches all core application assets, stylesheets, scripts, and specimen taxonomies into Cache Storage. Once opened once, you can take your device deep into a national park with zero internet connectivity.

Why Does Open Innovation Matter?

This project is the exact reason why open-source AI and local inference matter for humanity:

Cell Towers Don't Grow on Hiking Trails: Closed proprietary AI models (like OpenAI or Claude cloud endpoints) require continuous, high-bandwidth server roundtrips. When you hike into a valley, canyon, or national park with zero bars, closed AI is useless bricked software. Open-source local models are the only paradigm that democratizes intelligence offline.
Radical Location & Biometric Privacy: When people hike or forage for rare wild mushrooms (like Chanterelles or Morels), they do not want their coordinates and geotagged trail photos uploaded to centralized servers or scraped for advertising. With WildPulse, inference happens strictly on the user's silicon. No telemetry leaves the phone.
Zero Marginal Cost for Conservation & Education: Schools, state parks, and nature lovers shouldn't have to pay $0.03 per API token just to teach a kid how an oak leaf produces sugar. Open-weight innovation enables infinite, free exploration.

Taking It Outside: Field Testing on the Trail (Bonus Points!)

On Wednesday afternoon, I took WildPulse on a trail test in a local park. I switched my phone into Airplane Mode (Wi-Fi off, Mobile Data off) and plugged in my earbuds.

When I spotted an amber Sugar Maple leaf, I pointed my camera and did a 2-second capture:

Within 1 second, the audio guide spoke softly in my ear: "Sugar Maple (Acer saccharum). The brilliant orange and crimson hues indicate anthocyanin production, protecting the leaf tissues from light damage as chlorophyll breaks down before winter."
As soon as the voice started, Pocket Mode engaged: the screen went dim with a gentle green breathing pulse, reminding me to slip the phone into my jacket pocket.

Instead of spending 10 minutes squinting at a bright screen, I spent 15 seconds listening and the next 45 minutes walking with my head up, looking at the canopy and hearing real birds in the branches. It genuinely made the screen the shortest part of the outdoor experience.

My Agent Session

This project was built and recorded using DevRelay and the Antigravity agentic harness. You can inspect the full tool calls, prompt logs, and decision trees preserved in our normalized session:

WildPulse: Zero-Signal Trail Naturalist (Touch Grass)

You

Tested on the trail in airplane mode. The pocket mode audio played in my earbuds and I kept walking without looking at the screen.

Agent

▸ field_journal_logger Log Sugar Maple observation to local device journal.

Input

Log Sugar Maple observation to local device journal.

Output

Specimen logged locally: Acer saccharum. 100% private, zero telemetry transmitted.

Field observation saved! WildPulse successfully completed trail run with zero dropped frames and zero cellular signal.

(Direct Session Link: dev.to/agent_sessions/wildpulse-zero-signal-trail-naturalist-touch-grass-imku9p)

Prize Categories

Hacktoberfest Open-Source AI Challenge: Week 1 (Touch Grass) (Overall Main Prize)
Best Use of ElevenLabs (Integrated ElevenLabs Voice API for cinematic wildlife documentary narration)
Best Use of Render (render.yaml configured for zero-configuration static PWA deployment)

Now grab your boots, step outside, and touch some grass! 🌲🍂

BiteSize: The Open-Source Anti-Overwhelm Agent I Built for My Best Friend Mayank

Yash Raghubanshi — Fri, 02 Oct 2026 09:28:20 +0000

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

What I Built

Who I Built It For

I built BiteSize for my closest friend and roommate, Mayank.

Mayank is an exceptionally talented computer science senior. When he enters hyperfocus, he can construct distributed database prototypes and optimize complex algorithms without flinching. But like many brilliant neurodivergent minds with ADHD, Sunday evenings are a recurring nightmare.

When a messy week catches up to him—unwashed dishes, clothes strewn across the floor, broken lab code, unread messages from his parents, and impending assignment deadlines—his brain hits a physiological roadblock known as ADHD Executive Dysfunction Freeze.

To a neurotypical person, a to-do list is a series of simple tasks. To Mayank's ADHD brain, the list looks like an insurmountable, 400-pound boulder. The sheer cognitive friction causes panic, shame, and paralysis. He ends up stuck in bed for hours, staring at his phone, trapped in a spiral of guilt.

Every commercial to-do app he tried (Notion, Todoist, Apple Reminders) made it worse. They presented 30 red overdue badges, rigid hierarchies, and loud notifications that screamed: "Look how far behind you are!"

As Mayank put it to me:

"I don't need another app telling me I'm failing. I just need someone to tell me where to put my first foot."

What BiteSize Does

BiteSize is an autonomous, open-source AI agent built to de-escalate cognitive overload.

Instead of demanding structured inputs, BiteSize lets Mayank dump his raw, unfiltered, frantic stream-of-consciousness:

"My room is a disaster clothes on floor desk covered in cups I have to finish my ML lab report due tonight but my code is broken and I haven't eaten lunch and need to reply to mom's message and wash dishes before roommate comes home..."

The BiteSize agent ingests this panic dump and executes an autonomous de-escalation workflow:

Identifies the Core Friction Blocker: Detects what the brain is magnifying and isolates physiological needs (e.g. low blood sugar) before intellectual chores.
Decomposes the Mountain into "Atomic 2-Minute Steps": Strictly enforces that every sub-task takes ≤ 120 seconds and requires minimal physical friction (e.g. "Pick up 3 shirts and drop them into the hamper. Just 3. Don't fold anything.").
Anti-Paralysis Focus Mode (The Blindfold): Unlike standard to-do lists, BiteSize hides every other task. The user only sees ONE single 2-minute pebble at a time.
Virtual Body Doubler & Dopamine Engine: Accompanies the user with calm, non-judgmental presence, an embedded circular 2-minute visual countdown, synthesized calming Brown Noise, and celebratory micro-rewards.

Demo

Interactive Terminal UI (TUI): Fast, zero-distraction terminal mode powered by rich.
Responsive Web Dashboard: Clean dark interface featuring real-time Web Audio Brown Noise synthesis, circular countdown timers, and dopamine confetti.

Live Workflow Walkthrough:

Run python3 cli.py --sample to see Mayank's messy backlog deconstructed in real-time.
Or start the dashboard with python3 cli.py --web to experience the single-step focus mode and ambient noise.

Code

The complete source code is open source and hosted on GitHub:

Yash990-bit / bite_size

🌱 BiteSize — Open-Source ADHD Task De-Overwhelmer Agent

Built for Hacktoberfest Weekend Challenge: Build for a Friend
Dedicated with love to Mayank 💚

BiteSize is an autonomous, open-source AI agent built to conquer ADHD paralysis and executive dysfunction. When an overwhelmed brain perceives a mountain of tasks, BiteSize ingests unstructured, emotional panic dumps, analyzes cognitive friction, and shreds the mountain into microscopic, unambiguous 2-minute atomic micro-steps—presenting them one by one in an anti-paralysis focus mode with zero guilt.

💡 The Story: Why I Built This for Mayank

Mayank is my closest friend and roommate. He is a brilliant computer science student who can write high-performance distributed systems in a flow state. But when faced with messy Sunday backlogs—chores, unread messages, broken lab assignments, and chores—his brain hits executive dysfunction freeze.

To an ADHD brain, a to-do list doesn't look like steps; it feels like an insurmountable 400-pound…

View on GitHub

Core Repository Structure:

bitesize/agent.py: Autonomous agent harness orchestrating perception, reasoning, and tool calls.
bitesize/llm_client.py: Open-weight model adapter (Ollama with Llama 3.2 / Mistral, plus an embedded offline fallback engine).
bitesize/tools.py: Tool definitions (TaskDecomposer, CognitiveFrictionAnalyzer, BodyDoubler, DopamineTracker).
bitesize/tui.py: Rich terminal user interface for keyboard-driven focus.
api.py: FastAPI server bridging the Python agent with modern client interfaces.
web/: Vanilla HTML/CSS/JS frontend with Web Audio API brown noise generator.
devrelay_export.py: Normalized agent transcript generator for DevRelay.

How I Built It

The Open-Source AI Stack

BiteSize is engineered from the ground up around open-source AI:

Open-Weight Foundation Models:
- Primary: Meta Llama 3.2 (3B Instruct) and Mistral (7B) served via Ollama.
- These models excel at structured JSON formatting, zero-shot task decomposition, and empathetic role-playing while fitting comfortably in 4GB–8GB of RAM on everyday consumer laptops.
Autonomous Agent Harness:
- Built in Python 3 using a structured loop: $$\text{Perception (Panic Dump)} \rightarrow \text{Cognitive Analysis} \rightarrow \text{Tool Decomposition} \rightarrow \text{Focus Grounding}$$
- Strict Pydantic schemas validate that the AI output adheres to neurodivergent-safe constraints (e.g. estimated_seconds <= 120).
Multi-Domain De-escalation Tools:
- TaskDecomposerTool: Slices high-friction tasks into concrete physical movements.
- BodyDoublingCompanionTool: Uses open-weight prompting to provide psychological safety without patronizing toxic positivity.
- DopamineTrackerTool: Calculates streak momentum and reward milestones.
- SensoryReliefTool: Web Audio API noise synthesis generating calibrated Brownian noise (1/f² spectral density) to suppress sensory distractions.
Resilient Local Fallback Engine:
- If a friend doesn't have an Ollama daemon installed or is running on a low-spec battery-saver machine, BiteSize includes an embedded deterministic semantic engine that operates with zero delay and 0 MB RAM footprint.

Why Does Open Innovation Matter?

Open innovation wasn't just a design preference for BiteSize; it was an ethical and practical necessity.

1. Radical Privacy for Vulnerable Mental States

A brain-dump is not standard work text. When someone is in an executive dysfunction spiral, their brain dump contains deeply personal, vulnerable confessions: unfiled taxes, medical appointments they've avoided for months, academic failure fears, hygiene struggles, and family friction.

Sending that raw data to closed corporate APIs (OpenAI, Anthropic) means transmitting a person's deepest vulnerabilities to remote third-party servers where it could be logged, reviewed, or used for model retraining. With open weights running locally via Ollama, Mayank's thoughts never leave his device. Period.

2. Airplane-Mode Reliability (Zero Distractions)

When someone with ADHD is overwhelmed, their biggest enemy is the internet itself. Opening a browser to an API-dependent cloud tool invites Slack pings, YouTube notifications, and instant rabbit holes. BiteSize runs 100% offline. Mayank can disconnect Wi-Fi, put his laptop in Airplane Mode, and get unblocked without digital noise.

3. Zero Cost & Eternal Access

Commercial mental health and ADHD tools are infamous for putting basic productivity behind $15–$30/month subscription paywalls. Students and people battling burnout cannot afford recurring subscription guilt. Because BiteSize is built on open-weight models, it is free forever, with zero token fees.

4. Custom Tuning for Neurodivergence

Closed models often refuse to be concise or insist on preachy, corporate boilerplate ("As an AI assistant, I recommend you make a to-do list..."). With open weights, we can tune system prompts, control temperature, adjust logits, or fine-tune small models specifically on ADHD cognitive workflows.

Handover & Mayank's Reaction (Bonus Points!)

On Sunday afternoon, I sat next to Mayank, pulled up BiteSize on his laptop, and told him to pour his messy panic into the prompt.

He dumped 15 frantic sentences about his messy room, his broken ML lab, and his missed messages. BiteSize analyzed the friction, synthesized a calming background hum, and presented only Step 1:

"Pick up 3 pieces of clothing from the floor and drop them into the hamper. Just 3. Don't fold anything."

Timer: 90 seconds.

Here is what Mayank said verbatim:

"Bro, usually looking at my to-do list makes me want to close my laptop and go back to sleep. Seeing just 'Put 3 shirts in the basket' with a 2-minute countdown actually made me get off the bed. This is the first time an app didn't make me feel guilty."

Within 40 minutes, he had cleared his room, drunk water, eaten lunch, and written the first test case for his lab report. The paralysis was broken.

My Agent Session

This project was built and recorded using DevRelay and the Antigravity agentic harness. You can inspect the full tool calls, prompt logs, and decision trees preserved in our normalized session:

BiteSize: De-escalating ADHD Task Paralysis for Mayank

You

My room is a complete disaster clothes on floor desk covered in cups. I have to finish my machine learning lab report due tonight but my code is broken, and I haven't eaten lunch and need to reply to mom's message and wash dishes before roommate comes home. I feel completely frozen and can't get out of bed.

Agent

Analyzing cognitive friction and task ambiguity for Mayank. Breaking down executive paralysis...

▸ cognitive_friction_analyzer Unstructured panic dump across 4 domains (Physical, Academic, Admin, Self-Care)

Input

Unstructured panic dump across 4 domains (Physical, Academic, Admin, Self-Care)

Output

Identified core blocker: Overwhelming perceived mountain with severe task initiation resistance.

▸ task_decomposer Shred mountain into sub-120-second atomic micro-actions with zero guilt.

Input

Shred mountain into sub-120-second atomic micro-actions with zero guilt.

Output

Generated 5 atomic steps: (1) Drink glass of water, (2) Pick up 3 shirts, (3) Clear 2 cups from desk, (4) Open lab file, (5) Send 1-sentence reply to mom.

▸ body_doubling_companion Step 1 of 5 initiation

Input

Step 1 of 5 initiation

Output

'Hey Mayank, take a breath. You don't have to fix your whole life today. Just give me 60 seconds on this single glass of water. I am keeping time for you.'

Anti-Paralysis Focus Mode engaged. Presenting Step 1 with a 60-second micro-timer and Brown noise ambient sound.

You

Drank the water. Ready for the next tiny thing.

Agent

▸ dopamine_tracker Register completion of step-selfcare

Input

Register completion of step-selfcare

Output

Streak: 1 🔥 | Dopamine Points: +15 | Milestone: '🚀 Inertia Broken! The hardest part is behind you.'

Incredible job! Next atomic step: Pick up 3 pieces of clothing from the floor and drop them into the hamper. Only 3.

(Direct Session Link: dev.to/agent_sessions/bitesize-de-escalating-adhd-task-paralysis-for-mayank-xuvz13)

Prize Categories

Hacktoberfest Weekend Challenge: Build for a Friend (Primary)
Open-Source AI & Local Inference

Built with 💚 for Mayank, and for anyone who has ever felt paralyzed by their own to-do list.

I Built Free GenAI & ML Notes for Beginners (Hinglish +English+ Practical)

Yash Raghubanshi — Tue, 07 Apr 2026 19:06:40 +0000

🚀 I Created Free GenAI & ML Notes for Beginners (Hinglish + English+Practical)

Hey developers 👋

I’ve been learning Machine Learning and Generative AI, and one problem I faced was:

👉 Too much scattered content

👉 Too much theory, less clarity

👉 Hard to revise concepts

So I decided to build something simple 👇

📚 What I Built

I created a structured GitHub repository of:

🔹 Machine Learning basics

🔹 Generative AI concepts

🔹 Easy explanations (Hinglish style)

🔹 Practical intuition (not just theory)

👉 Check it out here:

https://github.com/Yash990-bit/Gen-AI-ML-notes

🧠 What You’ll Learn

This repo covers:

Linear Regression, Ridge, Lasso
EDA (Exploratory Data Analysis)
Probability & Statistics basics
ML concepts simplified
Generative AI fundamentals
Real-world intuition

💡 Why This is Different

Most resources are either:

❌ Too theoretical

❌ Too complex for beginners

This repo focuses on:

✅ Simple explanations

✅ Real-life analogies

✅ Easy revision notes

✅ Beginner-friendly learning path

🎯 Who is this for?

Beginners in ML/AI
Students preparing for interviews
Anyone who wants quick revision notes
Developers moving into AI

🔥 My Goal

To make AI/ML learning:

👉 Simple

👉 Practical

👉 Accessible to everyone

🤝 Contributions Welcome

If you want to:

Add topics
Improve explanations
Share examples

Feel free to contribute!

📢 If you found it useful

⭐ Star the repo

🔁 Share with others

💬 Give feedback

🚀 Let’s Learn AI Together

We don’t need 100 resources.

We need 1 clear resource.

👉 That’s what I’m trying to build.