📮 PhilatelyAI: An Offline, Zero-Cost AI Appraiser for High-Value Stamp Collections

Yash Mathur — Thu, 15 Oct 2026 12:06:48 +0000

📮 PhilatelyAI: An Offline, Zero-Cost AI Appraiser for High-Value Stamp Collectors

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸

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

What I Built

I built PhilatelyAI, a 100% offline, privacy-first desktop application that uses open-weight vision models to instantly appraise stamps, extract multilingual text/postmarks, and generate SEO-optimized marketplace listing drafts for platforms like eBay and Delcampe.

Who Would Pay For It?

High-end stamp collectors, dealers, and auction houses sit on thousands of unlisted, uncataloged assets. Digitizing these collections for insurance or resale is incredibly tedious.

More importantly, they refuse to upload high-resolution scans of valuable, unlisted assets to untrusted cloud APIs due to strict privacy and security concerns. They don't want their undocumented $5,000 stamps sitting on a third-party server.

Dealers would easily pay a $49 one-time license for a tool that automates their cataloging while keeping their data 100% offline and air-gapped.

The "First User" Bonus

I built this specifically for my grandfather (nanaji), who is a member of the Philatelic Collectors of Ahmedabad and owns a massive, documented mint-condition stamp collection. I sat down with him and ran a rare 2019 "Ahimsa Parmo Dharma" Mahatma Gandhi stamp through the local app.

His reaction? He was shocked. The local model not only identified the English text, but accurately extracted the microscopic Hindi script ("अहिंसा परमो धर्मः") and correctly flagged it as a mint, uncancelled specimen. He immediately noted that the "Copy for Delcampe" button would save his collector group hundreds of hours of typing out manual listings.

Why Open Innovation Matters Here

This side hustle is only possible because of open-source AI.

If I used a closed-source vision API (like OpenAI or Anthropic):

The Privacy Dealbreaker: Sending scans of undocumented, high-value assets to the cloud violates the core trust of high-end collectors.
The Profit Killer: Processing a 5,000-stamp binder via API would cost the dealer real money in token usage.

Because PhilatelyAI runs entirely locally on my AMD RX 580 using an open-weight model:

Complete Privacy: The data never leaves the user's hard drive.
100% Profit Margin: My inference cost is exactly $0. Once the software is sold, I don't have to worry about recurring API bills eating into my margins.

Best Use of Gemma

General-purpose, lightweight models often struggle with the dense, multilingual typography found on vintage stamps. I utilized Google's officially supported multimodal Gemma 4 (gemma4:e4b) running natively via Ollama.

Despite being small enough to run smoothly on my 8GB AMD GPU, Gemma 4 effortlessly handled the complex vision tasks—parsing microscopic Hindi and English inscriptions, detecting postmark degradation, and formatting everything into a strict JSON payload while generating the SEO-optimized marketplace draft.

Demo & Code

GitHub Repo: [https://github.com/yashmathur1207/PhilatelyAI]

Tech Stack:

AI Engine: gemma4:e4b running via Ollama
Backend: FastAPI (Python)
Frontend: HTML/CSS/JS with a custom Claymorphism UI

Touch Grass Agent: The AI That Yells At You To Go Outside

Yash Mathur — Sun, 11 Oct 2026 01:51:27 +0000

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

What I Built
The Touch Grass Agent is a multi-agent developer burnout detection and intervention system. It actively monitors developer telemetry—such as session duration, keystrokes per minute, and error rates—to calculate real-time fatigue probability.

When a developer crosses the burnout threshold, the agent executes a "System Override." It bypasses standard notifications to deliver a strict, synthesized audio directive commanding the developer to step away from the keyboard. It doesn't just tell you to take a break; it pulls live weather data and local geography to assign a specific, actionable destination (e.g., "March to Meghdoot Garden. It is 26°C and partly cloudy, you have no excuse. Go touch grass!"). It is built specifically for hyper-focused developers, hackathon participants, and remote workers who lose track of time and screen hours.

Demo
Live Backend :
Live Dashboard:
(Note: The frontend communicates securely with the live FastAPI backend via CORS. Ensure your volume is up to hear the agent's override directive!)

Code
https://github.com/yashmathur1207/touch-grass-agent

How I Built It
The architecture relies on a specialized, decoupled stack designed for speed and localized data processing:

AI & Inference Ecosystem: I utilized TabPFN (an open-weight tabular foundation model) to process the raw numerical developer telemetry (duration, keystrokes, error spikes) and output highly accurate burnout probabilities without the latency of a massive text-based LLM. Inference is powered by the Nebius API.

Agentic Tools: The orchestrator utilizes a custom serp_tool to fetch real-time local park data and weather conditions, while an elevenlabs_tool generates the commanding .mp3 audio directives on the fly.

Backend: Built with FastAPI, fully integrated with Sentry SDK for transaction tracing (sentry_sdk.start_transaction(op="task", name="Evaluate Burnout")) to monitor API performance and error states. Hosted on a Render web service.

Frontend: A lightweight, dark-mode terminal UI built with pure HTML/JS and Tailwind CSS. It connects to the FastAPI backend, dynamically renders the intervention dashboard, and streams the static audio file directly to the browser.

Why Does Open Innovation Matter?
Open innovation was critical for this build because burnout detection shouldn't require sending thousands of telemetry data points to a closed, opaque LLM provider. By leveraging open-weight models like TabPFN, the agent can perform inference on numerical tabular data quickly, locally, and efficiently. Open innovation empowers developers to stitch together highly specialized, modular AI tools—tabular models for telemetry, search APIs for local grounding, and voice generation for the UI—into a single, cohesive workflow rather than relying on a one-size-fits-all closed box.

Prize Categories
Best Use of Render: The FastAPI multi-agent backend is hosted as a live Web Service on Render, acting as the primary AI runtime.

Best Use of TabPFN: Used TabPFN's tabular foundation model to process numerical developer telemetry (duration, keystrokes, errors) and accurately classify/predict the developer's burnout probability.

_Best Use of ElevenLabs: _Used to generate the .mp3 synthesized voice narration for the agent's overriding directives, directly streamed to the frontend dashboard.

Best Use of Sentry Agent Tracing: Integrated the Sentry SDK to trace the agent's core evaluation task (sentry_sdk.start_transaction(op="task", name="Evaluate Burnout")), monitoring API latency and error states.

Best Use of SerpApi: Utilized a custom SerpApi tool to give the agent live search capabilities, grounding the intervention directive in fresh, local weather and geographic data (e.g., finding Meghdoot Garden in Indore).

DEV Community: Yash Mathur