Sloth — the lazy task triager (local Gemma does your organizing)

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3

What I Built

Sloth is a tiny CLI that takes your messy brain-dump — half-formed tasks, notes, random thoughts, copy-pasted reminders — and hands back a short, prioritized action plan with time estimates and one "do this first" pick.

The pitch is one line: dump your brain, Sloth sorts it. You never organize; the sloth does. It was built for Week 3's theme, Slow and Steady: software that rewards laziness. October 20 is International Sloth Day, and honestly, the sloths have it figured out — conserve energy, move at your own pace, let something else do the work.

Who would pay for this? Freelancers and students drowning in scattered to-dos — the same people who pay $10/month for AI organizers. Sloth does the core job (turn chaos into a plan) for $0, forever, with their words never leaving their laptop.

Demo

41-second demo, real run, nothing staged — a messy brain-dump.txt goes in, Gemma (running locally) deduplicates it, drops the noise, ranks every task, and picks the one thing to do first. The closing screen shows the trace: latency, tokens, cost $0.00.

Watch the 41-second demo (mp4)

Try it yourself — Python 3.10+, stdlib only, no pip install:

git clone https://github.com/raju12345-raju/sloth
cd sloth
# one-time: install Ollama and pull Gemma
curl -fsSL https://ollama.com/install.sh | sh && ollama pull gemma3:1b
python sloth.py triage examples/brain-dump.txt

Code

https://github.com/raju12345-raju/sloth

Public repo, MIT license, 13 passing tests, CI on every push. Built entirely inside the Oct 19–25 window.

How I Built It

The open-source AI is the core, not decoration. Remove it and the project collapses to a keyword guesser — which is exactly what the fallback mode is, and it's labeled as such.

The pipeline: your text goes to Gemma 3 (1B) running locally via Ollama. A strict system prompt asks for JSON only: deduplicated tasks, priorities 1–5 by impact-per-effort, effort estimates, one sloth_pick, and up to 3 ignored non-tasks. Sloth parses the JSON defensively (tolerates markdown fences and chatter), validates the shape, normalizes 1B-model quirks (greetings leaking into the plan, placeholder text in ignored), retries once on degenerate output, and renders a terminal plan. Every run appends a trace — model, latency, prompt/completion tokens, cost — viewable with sloth trace.

Why local: a brain-dump is private. Notes about money, health, clients — that text should never travel to someone else's server. Running Gemma on-device means the automation works offline, costs nothing per run, and can never be taken away by a pricing change.

Honest engineering notes: a 1B model occasionally misreads a task, so the README says plainly to verify before acting on anything important. If Ollama isn't reachable, Sloth says so out loud and drops to a transparent heuristic fallback — keyword scoring, clearly labeled, never pretending to be the model. AI is a layer here, not a lifeline: the tool is never useless, but it's only smart when the model is present.

Why Does Open Innovation Matter?

Three concrete reasons, not slogans:

1. Privacy by architecture. A closed API could do this triage — but every brain-dump would travel to a data center. With an open-weight model on my own machine, the embarrassing, financial, and half-baked thoughts in my notes stay mine. Open made that a default instead of a premium tier.
2. Zero marginal cost. Each triage costs $0.00 in tokens. A metered API turns "dump your brain daily" from a habit into a budget decision. Open weights make the lazy workflow — the entire point of the theme — economically sane.
3. Swappability. Don't like Gemma 3 1B? Set SLOTH_MODEL to any Ollama model. The provider is a 30-line abstraction; fine-tuning, quantization, or a bigger model are all one env var away. Try negotiating that with a closed endpoint.

My Agent Session

This build was pair-driven with an AI coding assistant (Muse Spark) — it wrote the first drafts, I directed the architecture, and every line was tested before commit. No DevRelay replay to embed; the repo's commit history and this write-up are the record.

Prize Categories

• Best Use of Gemma — Gemma 3 (open-weight) does the actual reasoning, run locally via Ollama. It is the product, not a garnish.
• Best Use of Sentry Agent Tracing — every run emits a trace (latency, tokens in/out, cost), inspectable via sloth trace; traces are shown in the demo.

What I didn't do (honest limits)

• No fake benchmarks, no invented user counts. The trace numbers in the demo (29s, 457 in / 408 out, $0.00) are from a real run on real hardware.
• The 1B model is the floor, not the ceiling — bigger local models give better triage, and the README says so.
• No "AI will change your life" claims. It sorts a to-do list. That's enough.