Someday 將截圖整理成待審核事項，並匯出日曆、費用與閲讀清單

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

What I Built

Every smartphone and desktop user has a digital graveyard of screenshots:

• Event posters on Instagram, Twitter, and Reddit with lines like "This Friday at 7 PM".
• Receipts and invoices from business dinners, coffee runs, and online services.
• Articles and technical papers bookmarked to read later.
• Addresses, confirmation numbers, and flight details.

Screenshots are the lowest-friction capture mechanism ever invented. But while saving a screenshot takes a split second, acting on them is tedious, manual drudgery. You have to open Google Calendar, decode relative dates, copy venues, switch between apps, or manually type receipt figures into a spreadsheet. Because this work is repetitive, it gets postponed indefinitely.

Someday is built for the Slow and Steady ethos. It rewards laziness:

Capture it once. Let Someday do the tedious part. Review the results over coffee.

Someday monitors your screenshot directory in the background. When new screenshots appear, an open-weight vision-language model inspects the image, extracts the critical details, resolves relative dates into exact calendar timestamps, and logs everything in a calm, paper-inspired Morning Ledger. You don't have to chat with a bot or prompt an AI. You simply wake up, review the ledger, approve with one click, and export valid .ics calendar events, CSV expense sheets, or reading lists.

Demo

• 🌐 Live Interactive Showcase (Vercel): https://someday-ledger.vercel.app
• 🎬 YouTube Demo Video (AI Narration): https://youtu.be/d7m2kpll3Ew
• 💻 Local Desktop Application: http://localhost:3000 (Frontend) & http://127.0.0.1:8000 (FastAPI Local Backend)
• 🔍 Primary Visual Moment: The Inspector Lightbox connects the raw screenshot side-by-side with the model's verbatim textual evidence quote, enabling human-in-the-loop review and instant approval.
• 📂 Multi-Format Native Exports:
  • Standard RFC-5545 iCalendar (.ics) for Apple Calendar, Google Calendar, and Outlook.
  • Expense CSV for Excel, QuickBooks, and Expensify.
  • Reading List CSV / JSON for Notion, Obsidian, and Readwise.

┌────────────────────────────────────────────────────────┐
│               The Morning Ledger UI                    │
├────────────────────────────────────────────────────────┤
│ [📅 Event] Open Source AI Summit · Tomorrow 7 PM       │
│    Quote: "Tomorrow at 7 PM - 10 PM at Hacker Hub"     │
│    [Approve] -> Exports to Calendar (.ics)             │
├────────────────────────────────────────────────────────┤
│ [🧾 Receipt] Blue Bottle Coffee · $17.09 USD           │
│    Quote: "TOTAL AMOUNT PAID $17.09 USD"               │
│    [Approve] -> Exports to Expense CSV                 │
├────────────────────────────────────────────────────────┤
│ [📖 Reading] How To Architect Reliable Local AI        │
│    Quote: "The Pragmatic Engineer · Gergely Orosz"     │
│    [Approve] -> Exports to Reading List JSON           │
└────────────────────────────────────────────────────────┘

Code

Shreyansh00987 / Someday

Your screenshots are a to-do list. Someday finishes them. Built for Hacktoberfest Slow and Steady Open-Source AI Challenge.

Someday

Your screenshots are a to-do list. Someday finishes them.

🌐 Live Vercel Interactive Showcase: https://someday-ledger.vercel.app
🎬 1-Minute Demo Video (AI Voiceover): Watch Demo Video
🖥️ Local-First Privacy Application: http://localhost:3000 (Next.js) + http://127.0.0.1:8000 (FastAPI + Ollama)

Someday is an open-source, local-first personal AI utility that turns forgotten screenshot folders into completed tasks. Instead of letting screenshots of event posters, dinner receipts, reading lists, and articles accumulate in digital oblivion, Someday monitors your screenshot folder, runs a local open-weight vision-language model, extracts structured actions with evidence grounding, and presents them in a calm morning ledger for your approval.

The Problem

Every day, people take screenshots of things they intend to do later:

• An event poster on Instagram or Twitter they want to attend.
• A paper receipt or dinner bill they need to expense.
• An article, paper, or book recommendation they want to read.
• An address or confirmation number they need to…

👉 GitHub Repository: https://github.com/Shreyansh00987/Someday

• Frontend: Next.js 16 (App Router), React 19, TypeScript, Tailwind CSS v4, Lucide Icons, and Impeccable Design System.
• Backend: Python 3.14, FastAPI, SQLite with WAL mode, Pydantic v2, Pillow, python-dateutil, icalendar.
• Inference Runtime: Local Ollama daemon running open-weight vision models (moondream 1.4B and gemma3:1b).
• Tests: 18 comprehensive automated tests covering unit math, database integrity, exports, API endpoints, and end-to-end acceptance.

How I Built It

Someday is architected around the principles of local-first privacy, deterministic guardrails, and asynchronous laziness:

1. Passive Folder Watcher (backend/watcher.py):

   • Implements debounced file size stability checks so partially written screenshots aren't read prematurely.
   • Computes SHA-256 hashes to guarantee duplicate screenshots are never re-processed.
   • Pillow header verification catches corrupt images safely without crashing the queue.
   • Preserves state across system restarts using SQLite.

2. Open-Weight Vision AI Pipeline (backend/inference.py):

   • Feeds screenshots directly to local Ollama (moondream:latest 1.4B with multimodal CLIP projector, and gemma3:1b).
   • Categorizes into event, receipt, reading, or uncertain.
   • Requires verbatim textual grounding: the model must quote exact text seen in the image.
   • Enforces Pydantic schema validation with an automated 1-shot repair retry.
   • Ambiguous or low-confidence screenshots are routed to the Needs Review queue rather than hallucinating false information.

3. Deterministic Date Resolution Engine (backend/date_engine.py):

   • Vision models are notoriously prone to math errors when calculating calendar dates. Someday eliminates this: the model only extracts the raw text expression ("Tomorrow at 7 PM", "This Saturday").
   • Our deterministic Python engine anchors the relative expression against the screenshot's file creation timestamp and outputs timezone-aware RFC-3339 / ISO-8601 datetimes.

4. The Morning Ledger Frontend (frontend/src/):

   • Crafted using the Impeccable Design System: warm tactile paper surfaces (#F7F5EE), deep charcoal ink typography (#181613), British racing green accents (#1E3F2E), and polaroid contact-sheet thumbnails.
   • Includes bulk approval, real-time client search, and dynamic WebP thumbnail caching.

Why Does Open Innovation Matter?

Open innovation is not just an implementation detail for Someday — it is the only reason the product is viable:

1. Zero Data Leakage for Sensitive Personal Information:
   Screenshots frequently contain banking balances, home addresses, private event invites, and medical appointment slips. Sending these to third-party proprietary cloud APIs is an unacceptable privacy hazard. Running open-weight models locally ensures not a single byte leaves the user's machine.

2. True Offline Independence:
   Once installed, Someday works on an airplane, in a remote cabin, or during an internet outage. It requires zero active internet connection to transcribe and organize your day.

3. Freedom from Per-Token Tolls:
   Proprietary vision APIs charge steep per-image fees. A backlog of 1,000 screenshots could cost tens of dollars. With open-weight models running on local hardware (our tests ran on a consumer laptop NVIDIA RTX 4050 GPU with 6GB VRAM), processing 10,000 screenshots costs $0.00 in API fees.

4. Model Substitutability:
   Users are never locked into a single provider. In Someday's settings, you can swap between moondream, gemma3:1b, minicpm-v, or llava with a single click.

The Lazy Payoff

We built an automated evaluation harness (backend/eval_harness.py) and benchmarked the workflow against real screenshot test fixtures:

• Real Measured Latency: ~10.58 seconds / screenshot on an NVIDIA RTX 4050 Laptop GPU.
• Memory Footprint: ~1.2 GB VRAM.
• Corrupt Image Resilience: 15.0 ms fail-safe recovery with zero queue interruption.
• Duplicate Overhead: 0 ms (instant SHA-256 cache hit).

In International Sloth Day spirit: Capture once, do nothing, review over coffee.

Prize Categories

Featured Categories:

• Best Use of Gemma: Google Gemma open-weight architecture (gemma3:1b) verified, installed, and integrated into our local Ollama registry for structured classification and text reasoning.
• Top Prize (Slow and Steady): Built specifically for Hacktoberfest Week 3 to automate the tedious backlog of digital clutter with quiet, background, local-first intelligence.