DEV Community: Abhinav Nayak

LazyMorning — Your Overnight Life Assistant

Abhinav Nayak — Mon, 26 Oct 2026 05:31:39 +0000

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

What I Built

Mornings are often ruined by a barrage of questions: What was I working on yesterday? Which deadlines are due today? What do I need to tackle first? This morning decision fatigue drains our best mental energy before the day even begins.

LazyMorning is an overnight personal life assistant designed to let you wake up with complete clarity and zero panic. While you sleep, LazyMorning analyzes your active tasks, captured notes, upcoming deadlines, and approved documents to prepare a serene, focused Morning Briefing delivered right as you start your day.

What it does:

Today's Focus: Synthesizes a calming, single-sentence priority focus for the morning.
Top 3 Priorities: Extracts the three highest-impact tasks with explicit reasoning and urgency tags.
Upcoming Deadlines: Clear timeline of impending deliverables so nothing slips through the cracks.
Unfinished Items: Quick tally of pending items requiring future attention.
Suggested First Step: Gives you an immediate, low-friction, concrete first action (e.g. "Spend 10 minutes reviewing the first slide") to kick off your day with calm momentum instead of paralysis.

Who gets to be a little lazier because of it?

Developers, students, founders, and busy professionals who want to reclaim peaceful mornings. Instead of spending 30 minutes every morning triaging sticky notes, Slack messages, and todo lists, you open your LazyMorning dashboard, sip your coffee with our relaxed sloth mascot, and know exactly what to do next.

Demo

🌐 Live Deployed App: https://lazymorning.onrender.com/
📖 Interactive API Documentation (Swagger): https://lazymorning.onrender.com/docs

Code

Check out the full open-source repository on GitHub:
👉 https://github.com/abhinavnayak21/LazyMorning-x-MLH

abhinavnayak21 / LazyMorning-x-MLH

🦥 LazyMorning — Your Overnight Life Assistant

"Slow plans, brighter days."
Built for the "Slow and Steady" Open-Source AI Challenge.

LazyMorning is a privacy-first, 100% local personal morning assistant. While you sleep, LazyMorning reviews your active tasks, notes, deadlines, and approved files to prepare a serene, focused morning briefing—eliminating morning decision fatigue and friction.

📸 Design & Experience

Crafted to match modern, calming aesthetics with an adorable mascot and clean cards:

Dashboard: Daily greeting, Today's Focus hero banner, 4 vital metrics (Top Priorities, Upcoming Deadlines, Pending Tasks, Total Notes), latest briefing summary, and quick actions.
Morning Briefing: Today's Focus, Top 3 Priorities (with reasoning and deadlines), Upcoming Deadlines list, Unfinished Items, and a gentle Suggested First Step.
Notes: Create, organize, tag, and search thoughts and ideas.
Tasks: Track priorities (High, Medium, Low), due dates, and instant checkbox completion.
Imported Files: Scan and import local .txt…

View on GitHub

How I Built It

LazyMorning is designed from the ground up around open-source AI and private local-first computing:

1. Open-Source AI Architecture

Model: Qwen 2.5 (7B Instruct) — a state-of-the-art open-weight model renowned for structured reasoning, nuance, and concise summarization.
Hybrid Inference Pipeline:
- Local Ollama: Runs 100% locally on personal machines via Ollama (qwen2.5:7b).
- Cloudflare Tunnel (cloudflared): Allows remote deployment on Render to securely reach local Ollama instances over an encrypted Cloudflare Tunnel without public port forwarding.
- Cloudflare Workers AI: Cloudflare’s serverless GPU edge option running @cf/qwen/qwen2.5-7b-instruct for 24/7 cloud availability.
- Zero-Crash Local Fallback Engine: If Ollama or the network is ever unreachable, an intelligent rule-based synthesizer steps in automatically, ensuring you never miss a briefing.
Hallucination Validator: AI outputs are strictly cross-referenced against the SQLite database to ensure the model never invents ghost tasks or fake deadlines.

2. Tech Stack

Frontend: Streamlit styled with custom modern CSS and base64 vector art to create a warm, calming aesthetic with our sloth mascot.
Backend: FastAPI with async REST routing and OpenAPI documentation.
Database: SQLite with SQLAlchemy ORM for lightweight, durable local persistence.
Overnight Scheduler: APScheduler running in-process for automated morning digest generation at configurable local wake-up times.
Automated Tests: Comprehensive Pytest suite covering CRUD operations, file security gates, AI fallback, and hallucination filters (100% pass rate).

3. Privacy & File Approval Gate

To prevent accidental data leakage, imported local .txt and .md files enter a pending state. Only files explicitly toggled to "✅ Approved for AI" are included in the morning briefing context.

Why Does Open Innovation Matter?

Your daily notes, personal reflections, project roadmaps, and client meeting takeaways are sensitive personal data. Sending them to closed, proprietary cloud AI APIs:

Risks privacy: Exposes private thoughts and proprietary work to third-party data collection and training pipelines.
Introduces toll booths: Closed APIs charge per-token fees, making automated overnight background cron jobs expensive over time.
Creates fragility: Closed services suffer outages, policy changes, and sudden price hikes.

Open innovation changes the game. By using open-weight models like Qwen 2.5 running locally via Ollama or over self-controlled Cloudflare tunnels, LazyMorning demonstrates that users don't need to sacrifice their privacy or pay monthly API subscriptions to get intelligent, automated life assistance.

My Agent Session

Built with pair-programming assistance from Antigravity. You can view or replay the session here:

The Lazy Payoff

Saved 20–30 minutes every morning: Eliminates the cognitive load of opening 5 different tools (Notes app, Todoist, Markdown files, Slack reminders) just to figure out what to do first.
Zero Morning Panic: Instead of waking up to an overwhelming wall of 20 unorganized tasks, you get Top 3 Priorities and a gentle Suggested First Step.
Effortless Overnight Automation: The in-process APScheduler generates everything before you open your eyes. You literally wake up to finished organization.

Prize Categories

Slow and Steady Grand Prize
Best Use of Open-Weight Models (Qwen 2.5 & Ollama)
Best Local-First / Privacy-Preserving AI Application
Most Creative / Fun UI & Mascot Design

BillCraft AI: Invoice Generator + AI-Powered Invoice Reader

Abhinav Nayak — Wed, 14 Oct 2026 23:47:23 +0000

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

What I Built

BillCraft AI is a privacy-first, intelligent invoice studio built for freelancers, contractors, and small business owners. It automates the tedious work of creating invoices and digitizing messy receipts without sending private financial data to third-party cloud servers.

The Problem It Solves

Freelancers and solopreneurs spend hours every month manually deciphering paper receipts, supplier invoices, and re-typing numbers into spreadsheets. Mainstream SaaS accounting tools charge costly monthly subscriptions ($30–$70/mo) and upload sensitive financial documents, client names, and banking details to proprietary cloud APIs.

The Solution

BillCraft AI solves this through two streamlined workflows:

Smart Invoice Generator: A fast workspace with persistent business profiles saved in local storage, dynamic real-time totals, one-click AI item description enhancement, and pixel-perfect branded PDF generation.
AI Invoice & Receipt Reader: A drag-and-drop document scanner that reads any PDF or image receipt using local OCR (RapidOCR + PyMuPDF), extracts structured line items via open-weight Qwen 2.5 7B, validates math integrity on the server, and transfers extracted data into the generator in a single click.

Freelancers, independent consultants, and micro-retailers who value financial confidentiality and want to save hours on bookkeeping would readily use and pay for this solution.

Demo

Live Web App: https://billcraft-wzr8.onrender.com
Local Ports: http://localhost:5173 (Frontend) | http://localhost:8000 (FastAPI Docs)

Code

abhinavnayak21 / BillCraft-ai-x-MLH

🧾 BillCraft AI

AI-powered invoice generator & reader built for freelancers and small businesses.
Powered by local Ollama (Qwen) for automated invoice data extraction and line-item enhancement, with strict server-side Python math validation and branded PDF export.

🚀 Key Features

Invoice Generator:
- Persistent inline business profile (Company Name, Email, Tax Rate %, Payment Terms) saved automatically in your browser's localStorage.
- Dynamic line items with auto-calculated subtotals, tax, and grand totals.
- ✨ AI Item Description: One-click enhancement of product/service descriptions using local Ollama.
- High-Resolution PDF Generation: Download professional branded PDF invoices matching the theme.
Invoice Reader (OCR + Local AI):
- Quick Drag & Drop: Drop any PDF document, scanned receipt, PNG, or JPG invoice.
- Document OCR: Uses RapidOCR and PyMuPDF for fast vector and scanned document text extraction.
- Local Ollama Extraction: Sends text to your local Qwen model (qwen2.5:7b) to extract vendor, client…

View on GitHub

GitHub Repository: https://github.com/abhinavnayak21/BillCraft-ai-x-MLH

How I Built It

BillCraft AI is designed around open-weight models and local inference:

Open-Weight Model: Qwen 2.5 7B (qwen2.5:7b) by Alibaba Cloud — selected for its high benchmark performance in structured JSON extraction and financial reasoning.
Local Inference Engine: Ollama running locally on localhost:11434, ensuring 100% private, zero-latency inference.
Cloud Fallback: For live demonstrations, the backend connects to Cloudflare Workers AI (@cf/qwen/qwen2.5-7b-instruct) for serverless, zero-cold-start inference on global GPUs.
OCR & Document Extraction: RapidOCR (ONNX runtime) combined with PyMuPDF (fitz) for spatial bounding-box layout clustering.
Math Verification Service: Strict server-side Python calculations reconcile line-item arithmetic (Quantity × Rate = Amount, Subtotal + Tax = Grand Total) before auto-saving into SQLite.
Tech Stack: React 19, TypeScript, Vite, Tailwind CSS v4, FastAPI, SQLAlchemy, SQLite, and ReportLab.

Why Does Open Innovation Matter?

Absolute Data Privacy: Invoices contain sensitive client information, bank details, and pricing margins. Closed APIs process and retain this data on their servers; running open-weight Qwen 2.5 on Ollama guarantees financial data never leaves the user's device.
Zero Marginal Cost: Closed AI APIs charge per token, making bulk invoice parsing cost-prohibitive. Open weights eliminate per-token charges entirely.
Offline Reliability: Small businesses and contractors often work in spots with spotty internet connectivity. BillCraft AI operates fully offline.
No Vendor Lock-In: Independent developers and businesses retain complete ownership over their infrastructure without fearing API price increases or unexpected deprecations.

My Agent Session

Throughout the development of BillCraft AI, I collaborated with Antigravity AI (Google DeepMind's agentic coding assistant) as an autonomous pair-programmer to:

Architect the multi-stage OCR and Qwen 2.5 JSON parser with server-side validation.
Build the full-stack UI using React 19 and Tailwind CSS v4.
Streamline deployment to Render with static file integration and Cloudflare Workers AI edge fallback.

Prize Categories

Open Source AI Champion
Qwen 2.5 7B locally via Ollama
Best Productivity / Developer Tool

SafeAir Walk: Find Your Clean-Air Window to Step Outside Today

Abhinav Nayak — Sat, 10 Oct 2026 17:38:17 +0000

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

What I Built

SafeAir Walk tells you the single best 1-hour window to step outside today — based on real air quality forecasts and heat — tailored to your health profile. Then a local open-source AI explains why in plain language, in your language of choice.

Air pollution and heat are invisible threats that people routinely ignore because the data is buried in dense charts. SafeAir Walk turns hourly AQI + apparent temperature forecasts into one clear, colour-coded verdict:

✅ Go for it — conditions are safe, head out now
🟡 Short walk only — keep it to 20–30 minutes
⛔ Skip today — air or heat is too high, try tomorrow

It is built for four health profiles: General, Child 👶, Elder 👴, and Asthma / Respiratory 💨 — each with tighter safety thresholds where it matters. The app uses your GPS or lets you search any city, then streams a warm, practical 2-sentence AI explanation — not a wall of warnings, just what to do and why.

The whole idea is to get people off the screen and safely into the world. When conditions are good, SafeAir Walk simply says "Go for it" and gives you a time. When they're bad, it suggests an indoor alternative and shows when tomorrow looks better.

No sign-ups. No API keys. No ads. Just open data and a fast, private Ollama LLM.

Demo

🌐 Live app: https://safeair-walk.onrender.com/

Search your city (or tap "Use my location") and you'll see:

The best hour to walk today
AQI, PM2.5, and feels-like temperature for that window
A streaming AI explanation personalised to your health profile powered by Ollama
An hourly dot-strip showing how the whole day looks at a glance

Code

abhinavnayak21 / safeair-walk

SafeAir Walk

Find your clean-air window to go outside. Checks AQI and heat for your area and picks the best 30–60 min window today.

Stack: React + Vite + Open-Meteo (no API key) + Ollama (local AI, optional)

Run locally

pnpm install
pnpm dev

Open http://localhost:5173

Enable AI explanations (optional)

# Windows PowerShell
$env:OLLAMA_ORIGINS="*"; ollama serve

The app auto-detects Ollama and uses it. Falls back to templates if offline.

Deploy (Free)

Two services:

HuggingFace Space → runs Ollama API (free Docker Space)
Render Static Site → hosts the React app (free)

Step 1 — Deploy Ollama to HuggingFace Spaces

Go to https://huggingface.co/new-space
Set:
- Space name: safeair-ollama
- SDK: Docker
- Visibility: Public
Upload the three files from the hf-space/ folder:
- Dockerfile
- hf_start.sh
- README.md
Wait ~5 min for the build + model download to finish
Your Ollama URL will be:
```
https://YOUR-HF-USERNAME-safeair-ollama.hf.space
```
Test it: curl https://YOUR-HF-USERNAME-safeair-ollama.hf.space/api/tags

Step

…

View on GitHub

How I Built It

Open-source AI: Ollama + Qwen 2.5 (Local Inference)

The AI layer runs entirely on Ollama with open-weight models — no external proprietary APIs, no OpenAI, no Anthropic.

Model: qwen2.5:1.5b / qwen2.5:7b — ultra-fast open-weight models capable of running seamlessly on consumer hardware.
Inference runtime: Ollama — models are served directly via ollama serve, exposing a streaming /api/generate and /api/tags endpoint with CORS enabled (OLLAMA_ORIGINS=*).
Streaming integration: The React frontend connects directly to Ollama's HTTP API, consuming NDJSON chunks via ReadableStream and TextDecoder to render tokens with a real-time typing effect.

The Scoring Engine & Deterministic Thresholds

To guarantee safety and scientific consistency, SafeAir Walk does not rely on LLM guesswork for health boundaries. All safety decisions are evaluated deterministically in src/safeair-core.js → PROFILES before any prompt is generated:

Profile	GO	SHORT	SKIP
General	AQI ≤ 100, Feels like ≤ 32°C	AQI ≤ 150, Feels like ≤ 36°C	Above thresholds
Child	AQI ≤ 75, Feels like ≤ 30°C	AQI ≤ 100, Feels like ≤ 34°C	Above thresholds
Elder	AQI ≤ 75, Feels like ≤ 30°C	AQI ≤ 100, Feels like ≤ 34°C	Above thresholds
Asthma	AQI ≤ 50, Feels like ≤ 32°C	AQI ≤ 100, Feels like ≤ 36°C	Above thresholds

Fetch — two parallel requests to Open-Meteo (free, no API key): air quality API for US AQI + PM2.5, and weather API for apparent temperature across 48 forecast hours.
Rate — each hour is evaluated against the selected profile's limits. If temperature is unsafe even when AQI is low (or vice versa), the app strictly selects the safer, conservative rating.
Pick — the algorithm evaluates today's daytime hours (5 AM–8 PM), selects the optimal window, and alerts the user if conditions deteriorate soon after.
AI Role — Ollama receives these structured verdicts to craft warm, supportive, plain-language advice without hallucinating risk limits.
Fallback — if Ollama is offline or unavailable, deterministic templates instantly respond with zero downtime.

Frontend & Deployment

React 19 + Vite — fast single-page app.
Vanilla CSS — dark glassmorphism theme with accessible contrast and micro-animations.
Render — static web app hosting with automated build pipelines (pnpm install && pnpm build).

Layer	Technology	Cost
AI Engine	Ollama (`qwen2.5`)	Free & Open Source
Web App Hosting	Render Static Site	Free
Weather & AQI	Open-Meteo API	Free, no key required
Geocoding	Open-Meteo Geocoding API	Free, no key required

Why Does Open Innovation Matter?

With a closed API, I would have had to pay per token, accept rate limits, and transmit every user's health profile and location data to an external provider's servers.

Open innovation with Ollama made three critical things possible:

Privacy by architecture. Ollama runs locally and independently. User queries, locations, and personal health profiles never leave the user's control. There are no tracking scripts, user accounts, or cloud telemetry pipelines.
Zero subscription / API fees. Because Ollama runs open-weight models, developers and users are completely free from monthly subscription tiers, credit card requirements, or surprise billing spikes. Anyone can clone the repo and run the entire AI stack for free.
Complete prompt & latency ownership. With Ollama, we have full control over sampling parameters (temperature, num_predict), token streaming, and response structure. There are no unexpected model deprecations, forced safety overrides, or arbitrary rate limit throttling.

Open innovation turned what would normally be an expensive, privacy-invasive cloud service into a fast, private, community-owned tool that gets people safely outdoors.

Prize Categories

Ollama — Open-weight model inference (qwen2.5) with streaming API served entirely via Ollama
Render — React frontend deployed as a Render Static Site

ReadEase: Building an Assistive Reading Studio for My Friend’s Mom with Dyslexia

Abhinav Nayak — Sun, 04 Oct 2026 23:44:24 +0000

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

What I Built

I built ReadEase for my friend’s mom, who has lived with dyslexia her entire life.

Whenever she tries to read online articles, long emails, recipes, or digital documents, letters often seem to flip, rotate, and crowd together on the screen. The harsh glare of standard white web pages causes severe eye fatigue and headaches within minutes. Over the years, she found herself having to ask family members to read things out loud for her, something that took away her reading independence and left her feeling frustrated.

Commercial assistive apps either felt overly complicated, were cluttered with ads, or required expensive monthly subscriptions.

I wanted to give her a calm, dignified, and effortless reading space. ReadEase is a web-based assistive reading studio designed to eliminate visual crowding, restore reading comfort, and give her back full autonomy over whatever she wants to read:

- Weighted Dyslexia-Friendly Fonts: Integrates OpenDyslexic (weighted bottoms that anchor letters to prevent them from flipping or swapping in the mind’s eye), Atkinson Hyperlegible (crafted by the Braille Institute for unmistakable character distinction), and Lexend.

- Soft Contrast Backdrops: Replaces harsh screen glare with soothing, optometrist-recommended paper tints (Warm Cream, Sage Green, and Soft Slate) that reduce visual stress and Scotopic Sensitivity (Irlen syndrome).

- Guided Line Focus Ruler: A movable reading ruler that highlights the active line and dims out peripheral distractions, preventing her from accidentally skipping lines.

- Bionic Fixation Reading: Automatically bolds the initial letters of each word, gently guiding her eyes across the sentence so her brain doesn't have to work overtime to recognize word boundaries.

- Synchronized Read-Along (Dual Sensory Immersion): High-quality, warm neural voice narration synchronized with millisecond-precision karaoke word highlighting—allowing her to hear and see words simultaneously without losing her place.
- One-Click Text Simplifier & Multilingual Support: Automatically rewrites convoluted sentences into clear, accessible language, and translates text into her preferred language whenever needed.

Demo

- 🌐 Live Application: https://readease.onrender.com

- 🛠️ Source Code github.com/abhinavnayak21/ReadEase

The Experience

- Zero Distractions: A distraction-free canvas where she can paste any recipe, letter, or news article.

- One-Touch Customization: Effortlessly tweak font size, letter spacing, and line height to match her comfort level.

- Human-like Audio: Calm, patient reading voices that she can slow down or speed up at will.

Code
ReadEase is fully open source under the MIT License:

abhinavnayak21 / ReadEase

👓 ReadEase - Assistive Typography & AI Reading Studio

ReadEase is an open-source assistive typography studio and reading workspace built to make digital reading effortless for everyone—including neurodivergent readers, individuals with dyslexia or ADHD, language learners, and people with visual sensitivities.

🌟 Core Features

Feature	Description	Why It Helps
📖 Dyslexia Mode	Activates the `OpenDyslexic` typeface with weighted baselines, generous line height, and custom letter tracking.	Prevents letter rotation/inversion and reduces visual crowding.
🎯 Bionic Reading	Dynamically bolds the initial syllables/fixations of words across sentences.	Guides saccadic eye movements, reducing cognitive fatigue and boosting reading speed.
📏 Interactive Reading Ruler	A cursor-following focus bar with gentle contrast styling.	Prevents accidental line skipping and aids readers with ADHD or focus difficulties.
🔤 Typography Engine	Custom selection of accessibility fonts (`OpenDyslexic`, `Roboto`, `Nunito`, `Comic Sans MS`, `Georgia`).	Allows readers to customize letterforms for maximum visual comfort.

…

View on GitHub

Repository Link: https://github.com/abhinavnayak21/ReadEase

How I Built It
To make the app truly accessible, every detail had to prioritize cognitive comfort over complex UI:

Fluid Typography & Visual Ergonomics: We built a custom CSS custom-property pipeline that applies letter weighting, expanded kerning, and line-height multipliers dynamically without layout shifts.
Cohere Command-R+ (Cognitive Simplification & Translation): For complex medical notices or dense news articles, we integrated Cohere’s command-r-plus model. With carefully crafted few-shot prompts, it deconstructs compound sentences into simpler clauses while preserving the core meaning and tone.
Millisecond-Accurate Neural Speech (ElevenLabs Alignment): Because dual-channel processing (visual + auditory) dramatically reduces dyslexic reading errors, we used ElevenLabs’ timestamped speech API (/with-timestamps). It returns character-level offsets that we map directly to word spans in the React DOM for real-time, synchronized highlighting.

Why Does Open Innovation Matter?

Accessibility should never be locked behind a paywall.
An estimated 700+ million people worldwide live with dyslexia. Many are adults, parents, and seniors who simply want to read daily emails or news without strain, yet modern assistive technology is increasingly commodified into costly $15–$30/month subscriptions.

Open innovation made ReadEase possible:

- Independence & Privacy: She can read private letters, medical records, or personal recipes knowing her data is never sold or retained for ad targeting.

- Built on Open Research: ReadEase stands on the shoulders of open-access fonts like OpenDyslexic and Atkinson Hyperlegible, which were gifted to the public domain by researchers and typographers.

- Extensibility: Anyone with a family member facing similar challenges can fork this repository, add their native language, or customize presets to their loved one's specific visual preferences.

My Agent Session
Built and refined using an agentic pairing workflow to transition the architecture into a clean, unified full-stack web application ready for 1-click cloud deployment.

Prize Categories

Best Project Built for a Friend
Accessibility / Assistive Tech
ElevenLabs
Render