Unison: Zero-Distraction Offline Notification Engine

Description and use

Unison is a local notification aggregator that removes noise, deduplicates messages using vector embeddings, and summarizes them into single‑sentence action items, freeing users from constant app switching.

Problem

Users suffer notification fatigue and context‑switching across multiple messaging platforms, emails, and system alerts, leading to cognitive overload and missed critical information.

Outcome

Unison reduced a full day of hundreds of notifications to 8‑10 actionable items, saving roughly an hour of active reading and eliminating anxiety from unread badges.

Workflow

• 1. Passively ingest alerts from WhatsApp, Android SMS, emails, and system feeds.
• 2. Generate text embeddings with FastEmbed (BAAI/bge-small-en-v1.5) and store them in LanceDB.
• 3. Semantically deduplicate overlapping alerts via vector similarity.
• 4. Send deduplicated context blocks to Qwen2.5‑3B‑Instruct served locally via Ollama.
• 5. Use Instructor and Pydantic v2 to enforce JSON schema output (priority and one‑sentence action item).
• 6. Render processed data into a unified single‑page dashboard with Streamlit.
• 7. User checks dashboard twice a day, receiving 8‑10 concise action items instead of hundreds of raw notifications.

Evidence

Unison is a 100% local, air‑gapped engine that intercepts SMS, WhatsApp, emails, and system alerts, stripping out the noise to build a single consolidated dashboard.

Source

The core intelligence runs on Qwen2.5-3B-Instruct served locally via Ollama.

Source

To prevent the LLM from being overwhelmed by rapid‑fire pings, the ingestion pipeline routes raw notifications through FastEmbed (using the BAAI/bge-small-en-v1.5 model) to generate text embeddings.

Source

The entire processed dataset is then rendered into a unified, single‑page dashboard using Streamlit, avoiding the bloat of a complex web framework while keeping the interface functional and clean.

Source

It successfully condensed a full day of digital chaos into 8 to 10 actual, single‑sentence action items.

Source