Slow and Steady：無須趕急或依賴雲端的本機開源 AI 工作空間

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

What I Built

Slow and Steady is a personal AI workspace that runs on the machine in front of you.

The idea is sloth-shaped on purpose. Most “AI assistants” ask you to open an account, paste private text into someone else’s API, and keep paying per token. That is a lot of fuss for a tool that is supposed to take work off your plate. This app does the opposite: install it once, talk to an open-source model on localhost, and keep the conversation in a SQLite file on disk.

It helps with the work people put off: explain a hard topic, tidy a paragraph, sketch a project, walk through a code snippet, or just sit with a calm chat window. The computer does the drafting. You stay in the chair.

There is a Demo Mode so you can use the whole interface before you download a model. Demo replies are scripted and labelled as scripted. They never pretend to be a language model, and they never silently replace a failed local request.

When you are ready, connect Ollama, pick an installed model such as qwen2.5:1.5b, and chat for real. Streaming, stop, regenerate, Markdown, code highlighting, and conversation history all work the same way in both modes.

What it does

• Chat with streaming, stop, regenerate, copy, and Markdown
• Keep history locally: create, rename, delete, continue, clear all
• Choose Demo Mode or any model already installed in Ollama
• Tune temperature, max tokens, system prompt, and streaming
• Show an honest status: Demo Mode, Local AI Ready, Model Loading, Generating, Model Not Found, Ollama Offline, Backend Unavailable

Local AI Ready is only shown after the backend has checked that Ollama answers and the selected model is installed.

Demo

Code

• GitHub: https://github.com/gabutersproject/slow-and-steady
• Run locally: see the README — open http://127.0.0.1:4747 after starting both processes

Quick path for judges who do not want to pull a model:

./scripts/start-backend.sh    # terminal 1
./scripts/start-frontend.sh   # terminal 2

Click Start Demo. Suggested prompts are on the empty chat. Replies are labelled scripted, not an AI model.

To use a real open-weight model (not automatic — you pull it):

ollama pull qwen2.5:1.5b

Then choose that model in the sidebar.

How I Built It

Two processes, both on loopback.

Browser (Vite + React)  →  FastAPI on 127.0.0.1:8747  →  Ollama on 127.0.0.1:11434
                                ↓
                         SQLite (chats + settings)

The UI never talks to a model. An abstract AIProvider sits behind the API with two implementations:

1. DemoProvider — a small set of scripted replies, streamed in chunks so the UI path is the same as a real model.
2. OllamaProvider — local HTTP only (/api/tags, /api/chat). No cloud fallback. No auto-download.

Stack:

Security choices that match the theme of “do less, leak less”:

• Bind the API to localhost; CORS only for the local Vite origin
• Validate temperature, token limits, message length, and model names
• One generation per conversation at a time (no interleaved history)
• Stop/cancel persists partial text as cancelled, never as a finished answer
• No shell from model output, no secret leakage in error payloads, no arbitrary file tools in chat

Tests run without a downloaded model. Last run in the build environment: 46 backend tests passed, 17 frontend tests passed. Live generation against a real Ollama instance was not verified in that environment because Ollama was not installed there; offline detection and the mocked provider suite were.

Why Does Open Innovation Matter?

The closed path for a chat app is easy: drop in a paid API key and ship. It is also the path that sends every draft, code snippet, and half-formed question to a server you do not run.

Slow and Steady is built so the open pieces are the product, not a badge:

• Your words stay here. History is a local SQLite file. The model, when connected, is a process on 127.0.0.1.
• No meter running. After npm install, pip install, and an optional ollama pull, chat does not call a cloud LLM. You are not renting intelligence by the token.
• You can swap the brain. Any model Ollama lists is selectable. The UI does not care whether it is Qwen, Llama, or something else you pulled yourself.
• It still works when you are tired. Demo Mode is the sloth door: explore the app at 11pm without waiting on a multi-gigabyte download. Connect a real model when you have the energy.

A closed API would have been faster to stub and worse at the job. The job is a private, unhurried assistant that does the typing so you do not have to.

Prize Categories

Hacktoberfest Open-Source AI Challenge: Week 3 (theme: Slow and Steady) — overall

No partner runtime is required for this project. It is meant to live on a laptop, not on a hosted GPU. I am not claiming Render, Gemma-as-a-service, DigitalOcean, or other partner categories unless those are added later with a real integration.