LocalScribe AI

SWAGAT SOURAV DHAR — Sat, 17 Oct 2026 17:36:17 +0000

I built LocalScribe AI: A private, local-first AI assistant for receipts and invoices.

Freelancers and small businesses shouldn’t have to spend hours organizing receipts and invoices. But sending sensitive financial documents to external AI services isn’t always desirable, and API costs can add up.

So I’m building LocalScribe AI, an AI-powered document organizer that runs locally.

What it does

Extracts information from receipts and invoices.
Categorizes expenses.
Flags uncertain fields for human review.
Organizes records into a searchable table.
Exports validated data to CSV.

Why open-source AI?

The model is a core part of the application, not just an API call hidden behind a user interface.

By using a compatible open-weight model with local inference, LocalScribe AI can process documents without relying on an external AI API for its core functionality.

That means greater control over sensitive data, no per-request AI API fees, and the flexibility to experiment with different models.

The trade-offs are real: local inference requires suitable hardware, and extraction accuracy still needs to be tested.

Who is it for?

Freelancers, consultants, and small businesses that need a simpler way to organize financial documents.

What I’m testing

Can a local AI workflow save someone enough time to become a product they’d actually pay for?

My next step is to put the prototype in front of a real user, measure the results, and improve it based on their feedback.

Built with: React, Python, FastAPI, Ollama, an open-weight language model, OCR, and SQLite.

I’ll share the working demo, source code, and real user feedback as the project develops.

Open innovation matters when it gives people more control over their tools, their data, and the economics of using AI.

Built an Offline Frost Predictor with TabPFN-v2 and Gemma 3 (No Internet, No API Keys)

SWAGAT SOURAV DHAR — Fri, 09 Oct 2026 16:23:42 +0000

Gardeners often need planting advice where cell service is unreliable or nonexistent. To solve this, I built FrostWise, a local-first gardening assistant designed for the Hacktoberfest Open-Source AI Challenge.

Architecture
TabPFN-v2 (Tabular Foundation Model): Evaluates local weather history (ERA5 minimum daily temperatures) to forecast first and last frost dates without relying on generic hardcoded zone charts.
Gemma 3 (via Ollama): Translates predicted freeze risks into practical planting schedules and crop protection guidance.

Everything runs entirely on your local machine—no remote telemetry, no external API calls, and zero ongoing costs.

📺 Video Walkthrough: https://youtu.be/7gHYehieEnQ

Feedback and contributions are welcome!

DEV Community: SWAGAT SOURAV DHAR

LocalScribe AI

Built an Offline Frost Predictor with TabPFN-v2 and Gemma 3 (No Internet, No API Keys)