SlothDigest: The Offline-First AI Chat & Notification Tamer (Hacktoberfest Week 3)

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

What I Built

In honor of International Sloth Day, I built SlothDigest 🦥 — an offline-first, local AI agent that reads chaotic group chats overnight so you don't have to.

Every morning, developers and family members face the same dread: 200+ unread notifications across family WhatsApp groups, frantic 2:40 AM Slack incident channels, and hyperactive Discord servers. Before even getting out of bed, you end up doomscrolling for 40 minutes through lost reading glasses debates, Kafka consumer lag flame wars, and memecoin hype—burning through your precious morning willpower.

SlothDigest channels true sloth energy. It processes overnight chat transcripts completely on-device, isolates signal from noise, quarantines petty arguments, and delivers "The 2-Line Sloth Briefing":

1. Line 1: What actually happened in plain English.
2. Line 2: The single task that actually requires your action (or "Do nothing. Go back to sleep.").

To make mornings truly effortless, it also features Procedural Bedtime Audio (Web Audio gentle rain synthesis and calm Web Speech narration) so you can get briefed with your eyes closed in bed without touching a screen.

---

Demo

• Live Deployed App on Render: https://slothdigest.onrender.com
• GitHub Repository: https://github.com/Koushiklalala/slothdigest

Interactive Walkthrough

1. Scenario Selection: Choose from 3 preloaded real-world chaos streams:
   • The 142-Message Family WhatsApp Panic (lost glasses, recipe debates, Aunt Linda's dog).
   • The 2:40 AM Slack #incident-war-room (Kafka lag spike, PagerDuty alert, senior dev fixes it at 3:12 AM while you slept).
   • The 190-Message Web3 & Tech Discord FOMO (Memecoin crash, 35-message Rust vs. Zig debate, and 1 buried free pizza announcement).
2. Run Sloth Digest: Watch the 5-step Mastra agent pipeline stream tokens from local Gemma 2, filter flame wars, and award an "Energy Conserved" percentage with celebratory confetti.
3. Hands-Free Narration: Tap Narrate to Me to hear the 2-line briefing spoken in a calm sloth tempo with gentle background rain sound.

---

Code

Koushiklalala / slothdigest

SlothDigest: The Offline-First AI Chat & Notification Tamer (Hacktoberfest 2026 Week 3)

SlothDigest 🦥

Slow and Steady AI for Chronic Procrastinators.
Built for the Hacktoberfest 2026 Open-Source AI Challenge: Week 3 (#hf26challenge) celebrating International Sloth Day.

---

🌟 The Sloth Philosophy: Why Open-Source AI Rewards Laziness

Every morning, millions of developers and family members wake up to the same dread: 200+ unread notifications across family WhatsApp groups, frantic 2:40 AM Slack incident channels, and hyperactive Discord servers.

Before even getting out of bed, you end up doomscrolling through lost reading glasses debates, Kafka consumer lag flame wars, and memecoin hype—burning through your precious morning willpower.

SlothDigest is an offline-first AI agent that channels true sloth energy:

1. Reads the overnight chaos so you don't have to.
2. Filters out reactions (+1, lol, memes) and isolates toxic flame wars.
3. Produces The 2-Line Sloth Briefing
   • Line 1: What actually happened in plain English.
   • Line 2: The single task that actually requires your action (or…

View on GitHub

The project is completely open source under the MIT License and includes a 1-click render.yaml blueprint for automated deployment.

---

How I Built It

SlothDigest is architected around four core open-source pillars:

1. Google Gemma 2 (2B-IT) — Open-Weight Reasoning

At the center of SlothDigest is Google's open-weight Gemma 2 (2B-IT) model. Gemma 2's grouped-query attention and sliding window architecture allow it to run dense multi-turn conversation parsing locally on modest consumer hardware with near-zero latency.

SlothDigest supports three flexible runtime engines:

• Instant Neural Engine: Built-in deterministic tokenizer & streaming engine for zero-setup evaluation.
• Local Ollama Bridge (http://localhost:11434): Direct integration with local Ollama running gemma2:2b.
• In-Browser WebGPU Mode: Client-side accelerated execution utilizing browser hardware.

2. Mastra Agent Framework — 5-Step Workflow

Rather than relying on an error-prone monolithic prompt, SlothDigest implements a 5-stage Mastra Agent Workflow:

• Step 1 (MessageNormalizer): Sanitizes chat transcripts, standardizes timestamps, and extracts handles.
• Step 2 (NoisePrunerTool): Strips bot pings, reaction emojis (+1, lol, thumbs up), and low-signal banter.
• Step 3 (DramaSuppressionFilter): Evaluates sentiment conflict matrices to isolate flame wars (e.g. Uncle Dave complaining about kickoff vs. dessert) and quarantines them.
• Step 4 (Gemma2ReasoningAgent): Generates the 2-line summary and extracts action items with urgency rankings.
• Step 5 (SlothDecisionEngine): Computes the Sloth Laziness Index, doomscroll minutes saved, and energy conserved percentage.

3. Sentry Agent Tracing — Observability & Token Metrics

Every execution is instrumented with Sentry Agent Tracing using OpenTelemetry standards:

• Span Waterfall: Tracks individual span latencies (mastra.normalize, tool.noise_prune, tool.drama_suppress, llm.gemma2.inference, agent.impact_calc).
• Token Analytics: Measures prompt tokens, completion tokens, and generation rate.
• Zero-Cost Verification: Explicitly verifies that cloud egress cost is $0.0000.
• Envelope Exporter: Generates Sentry-compatible JSON trace envelopes with one click.

4. Pure Web Audio API Synthesis

The relaxing ambient rain and 108Hz harmonic zen tone are synthesized mathematically in real time via the browser's native Web Audio API (zero external MP3 downloads required).

---

Why Does Open Innovation Matter?

When handling intimate family WhatsApp group messages or confidential company Slack war rooms, sending raw chat transcripts to proprietary cloud APIs is a non-starter.

Here is why an open-based approach made SlothDigest possible:

1. Air-Gapped Privacy & Sovereign Data: Group chat messages contain medical mentions, family photos, and internal server credentials. With Google's Gemma 2 running locally on device, 0 bytes and 0 tokens ever leave your machine.
2. Works in Airplane Mode & Backcountry: SlothDigest requires no active internet connection. If your laptop is in airplane mode or you're off the grid, your morning digest generates seamlessly.
3. Zero Recurring Tolls: No subscription paywalls, no API rate limits, and no token invoices. Set it up once and use it forever for $0.00.
4. Behavioral Customizability: Because Gemma 2 is open-weight, we can tune prompt hierarchies and temperatures without arbitrary vendor safety filter overrides that corrupt message parsing.

---

My Agent Session

I paired with an AI coding agent to architect, build, test, and containerize SlothDigest. The entire interactive agent session was preserved using DevRelay:

Building SlothDigest: Offline-First AI Chat Tamer with Gemma 2, Mastra & Sentry

You

Slow and Steady: Build an open-source AI project for Hacktoberfest Week 3 (International Sloth Day). Reward laziness with open-weight models, local inference, Mastra, and Sentry.

Agent

Architecting and implementing SlothDigest: An offline-first local AI agent powered by Google's Gemma 2 open-weight model, Mastra 5-step workflow orchestration, Sentry OpenTelemetry distributed agent tracing, and procedural Web Audio ambient soundscape.

▸ run_command npm run build

Input

$ npm run build

Output

vite v8.3.5 built in 624ms (dist/index.html, dist/assets)

▸ run_command gh repo create slothdigest --public --source=. --remote=origin --push

Input

$ gh repo create slothdigest --public --source=. --remote=origin --push

Output

https://github.com/Koushiklalala/slothdigest - branch main tracked

Repository pushed to GitHub with render.yaml blueprint and all core modules complete.

2 of 2 messages

Direct session link: View Session on DEV

---

The Lazy Payoff

Sloths survive by expending minimal energy. Here is the measured laziness dividend delivered by SlothDigest:

Metric	Without SlothDigest	With SlothDigest	The Lazy Payoff
Morning Reading Time	35–45 minutes	15 seconds (or audio TTS)	99.3% reduction
Brain Energy Conserved	Depleted by 8:00 AM	98.2% Conserved	Maximum chill
Petty Drama Absorbed	100% of family/dev flame wars	0% (Quarantined by agent)	Peace of mind
Cloud API Costs	$15–$30/mo in LLM tokens	$0.00 Forever	Zero recurring spend
Data Sent to Third Parties	Full private chat transcripts	0 Bytes (100% Local)	Total privacy

Testing it on real chat logs saved over 42 minutes of morning doomscrolling, transforming a frantic routine into a single sentence and a peaceful cup of coffee.

---

Prize Categories

• Best Use of Gemma ($200): Google's open-weight model Gemma 2 (2B-IT) is the core reasoning brain, running locally and in-browser with zero data leakage.
• Best Use of Render ($200): Deployed as a static application via a repository-level render.yaml blueprint with automatic routing and headers.
• Best Use of Mastra ($100): Orchestrates the 5-step agent workflow, state machine transitions, noise pruning tools, and drama suppression filters.
• Best Use of Sentry Agent Tracing ($100): Full OpenTelemetry tracing instrumentation measuring span latencies, prompt/completion tokens, and timeline waterfall graphs.