DEV Community: Yug dhamecha

jwrhbgtiear

Yug dhamecha — Sun, 25 Oct 2026 15:29:45 +0000

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

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

The Lazy Payoff

Prize Categories

LocalLens: A Private, Local-First Invoice and Receipt Sorter for Freelancers

Yug dhamecha — Sun, 18 Oct 2026 17:03:08 +0000

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

What I Built

I built LocalLens, a private invoice and receipt sorter designed for freelancers and small business owners.

What it does: It automatically ingests PDF invoices and receipt images, runs local OCR, extracts structured data (vendor name, date, amount, and category), and sorts them into organized folders while exporting clean CSV summaries for tax accounting.
Who would pay for it: Freelancers, independent consultants, and small business owners who want to automate their tedious tax-season bookkeeping without exposing sensitive client financial data to third-party cloud servers.

Demo

(Add a link to a Loom video, YouTube demo, or deployed interface here)

Code

(Omitted as requested)

How I Built It

LocalLens is built around a fully local, open-source AI architecture:

Local Inference via Ollama: Powered by open-weight models (llama3.2 / mistral) running locally on the machine via Ollama.
Document Extraction: Utilizes pytesseract and pdf2image for offline OCR on PDF and image-based receipts.
Application Framework: Built using Python and Streamlit for an interactive, responsive desktop-ready interface.

Why Does Open Innovation Matter?

Open innovation was non-negotiable for this project because:

Total Client Data Privacy: Freelancers handle strict NDAs and confidential financial records. Cloud-based APIs require uploading sensitive invoices to external servers, risking data leaks. Local execution keeps all files securely on the user's hard drive.
Zero Marginal Costs: SaaS invoice parsers charge high per-document fees. Running open-weight models locally means ₹0 per API call, allowing a sustainable indie business model.
Flexibility: Users can swap out local models depending on hardware limits (e.g., swapping a lighter 8B model for speed).

My First User

I tested LocalLens with a freelance graphic designer friend who processed around 30 unorganized PDF client receipts.

Feedback: It saved them roughly 45 minutes of manual sorting. They loved that their confidential client branding and financial totals never left their laptop.
Monetization: They confirmed they would gladly pay a one-time fee of ₹800–₹1,200 (\$10–\$15) to own the tool outright without recurring subscription lock-ins.

Prize Categories

Open-Source AI / Side Hustle

TrailEar: A local AI birding companion that turns birdsong into a journal

Yug dhamecha — Sat, 10 Oct 2026 13:19:55 +0000

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

What I Built

TrailEar is a command-line birding companion that helps you identify birds by listening to their calls. Record audio outdoors or analyze an existing recording, and TrailEar uses BirdNET to detect species and their confidence scores. It then uses a local AI model to turn the detections into a short birding journal entry. It’s for birders, hikers, and anyone curious about the wildlife around them.

Demo

Try it from the command line:


bash
trailear record --minutes 2 --place "Pune"

I turned my Dadi's voice memos into a recipe book, fully offline

Yug dhamecha — Mon, 05 Oct 2026 04:13:18 +0000

What I Built

My Dadi, Savitri Devi, has cooked for our family for decades and has never written a single recipe down. Ask her for measurements and you get "thoda sa namak, jab tak rang badal na jaye." When she's gone, those recipes go with her.

So I built Dadi's Recipe Book. She records voice memos in Hindi while she cooks, and the tool turns them into a clean, printable recipe book. It transcribes the audio, structures it into ingredients and steps, and flags every vague amount ("thoda", "andaaz se", "ek mutthi") so I can ask her about it later instead of guessing.

Demo

[Add a screenshot of a Hindi transcript turning into a structured recipe, or a short video of it running with the wifi off]

Code

<------------------------------------------------------------------>

How I Built It

Everything runs locally on my laptop, with no cloud API anywhere:

Speech-to-text: Whisper (medium) via faster-whisper, with the language set to Hindi (hi)
Recipe structuring: Qwen 2.5 (7B) running through Ollama
Export: a Python script that builds a PDF with her name, photo, and recipes

The LLM prompt is written to keep her own Hindi phrasing instead of rewriting it into generic recipe-speak, and to mark any vague quantity as "Dadi se poochna hai" rather than inventing a number.

What went wrong: [Write one real failure, e.g. "Whisper small mangled dish names like baingan bharta, so I switched to medium."]

Why Does Open Innovation Matter?

Her voice stays private. These are recordings of a family elder. They never leave my machine, and no server I don't control ever hears them.
It works offline. I ran it at her house with no internet, on a laptop, while she was cooking.
I could swap models when something broke. When [real example: Whisper misheard a dish name / the LLM over-corrected her style], I changed the model size and edited the prompt. With a closed API, I couldn't have touched what's underneath.
It costs nothing to run, so the book can keep growing with every recipe she shares.

The Handover

I gave Dadi the printed book on [date]. Her reaction: "[her actual words]"

What's Next

More recipes, a copy for each cousin, and voice search so anyone can ask "how did Dadi make kadhi?"

hello

Yug dhamecha — Sun, 04 Oct 2026 17:35:44 +0000

this is my first post!!