Loading projects…
Build with System One models
Projects worth building, step by step, and the apps and tools people have built with decision models.
Build with System One models
Projects worth building, step by step, and the apps and tools people have built with decision models.
Build with System One models
Projects worth building, step by step, and the apps and tools people have built with decision models.
27 project ideas · step by step
Routers, guardrails, scorers, agents and games: things a decision model does better than a language model. Each idea names the decision, the models that suit it and the steps to build it.
pip install systemonemodels, then systemone pull one of the models it lists.Beginner projects. An afternoon: one model, one question, a small script.
Route every incoming ticket to the right queue in milliseconds, and send the ones the model is unsure about to a person.
Ask a decision model "is this command safe to run?" before a coding or ops agent executes it, and block or ask for approval below a threshold.
A decision model reads the prompt and picks the smallest model likely to handle it, cutting LLM spend without hurting quality.
Score posts or comments against each rule of your community policy, auto-hide the clear violations, and queue the borderline ones for moderators.
Turn free-text call notes and emails into a 1–5 lead score with a probability, so the team calls the right people first.
Sort incoming email into invoices, customer questions, sales enquiries and noise with a model small enough to run without a GPU.
Decide what kind of document arrived (invoice, contract, ID, receipt) and extract the fields that matter, with a confidence for each.
Give every payment a calibrated probability of being suspicious, fine-tuned on your past investigations, and route the risky ones to analysts.
Detect what users want in Thai, Portuguese, Hindi or Nepali without translating first, and route them to the right flow.
Play Snake, Tetris or Connect Four with a decision model choosing each move from the legal ones, with its confidence shown live.
Decide, for each field on a form, whether to type a value, tick a box, click or skip, using a model small enough to run in the loop.
Choose a robot's next action (pick, skip, move, stop) from a text description of what its sensors report, on a small board without a GPU.
Take a labelled dataset from your own work, fine-tune Laya on it locally, prove it beats the base, and publish it on the registry.
Answer every request with a decision model first, send only the uncertain ones to an LLM, and measure how much time and money it saves.
Score every pull request for risk from its diff and description, and ask for a senior reviewer when it touches something dangerous.
Serve an open decision model to Claude, Cursor or any MCP client with noulxp mcp, so the small decisions an agent makes over and over come back as probabilities it can act on.
Before a retrieval runs, a decision model reads the question and picks the index that holds the answer, or decides none is needed, so the LLM reads less noise and you run fewer searches.
Score every answer your LLM gives against a rubric, from 1 to 5 with a probability for each level, and check grounding and format with yes/no questions, fast enough for every test case and every release.
Score every incoming alert from 1 to 5 for severity, from its message, its service and how often it fired, so the person on call is woken only for the alerts that need them.
Tag every new review with the themes it mentions, such as crashes, pricing, login or a missing feature, score its sentiment, and send the product team a weekly count instead of a spreadsheet to read.
Ask whether each email you open is a phishing attempt, in a browser extension or a mail add-in, with a small model running on the same machine, and show a warning with the probability before anyone clicks a link.
Check every return or refund request against your policy before an agent opens it, with a probability for each rule, so clear cases are ready in seconds and a person still approves every refund.
Route support messages written in Devanagari, Romanized Nepali or a mix with English to the right team, with a model that reads Nepali as people write it, and hand the unclear ones to a person.
Pre-label a dataset with a decision model and its probability, accept the confident labels after a spot check, and send only the uncertain rows to people, so labelling takes a fraction of the time.
Find the sentences in meeting notes or a transcript that commit someone to doing something, pick each one's owner from the attendee list, and post a task list when the meeting ends.
Turn a spoken command into one of your home's actions, such as lights, heating, music or a timer, on a Raspberry Pi with no cloud service, and ask again when the model is unsure.
Turn a fine-tuned checkpoint into a NoulXP package, prove it answers exactly as your own code does, serve it over HTTP on any machine, and publish it so its page shows NoulXP compatible.