See Sparsi in action across a range of production patterns. These examples demonstrate how to move from "prompt-only" automation to robust, deterministic-first workflows—drastically reducing token usage and eliminating hallucinations by using AI only where judgment is required.
AI reads a free-text ticket and decides its category; deterministic rules route it down the matching lane, those lanes run in parallel, and the results merge into one structured brief. The canonical classify-route-fan-out shape.
Three AI extractors read the recipe in parallel; a deterministic formula turns the counts and cook time into a difficulty score, which routes to targeted cooking advice. AI reads the prose, code does the maths.
Fetches a repository's README, runs five independent AI quality probes at once, then averages the scores deterministically and appends a coverage warning. Independent AI assessments combined by code.
AI reads messy live weather text into numbers and classifies the conditions; deterministic rules derive temperature band and wet/windy flags, then AI drafts outfit advice with a separate extreme-weather check.
Runs an AI check over every story title in parallel, filters the results deterministically, lets the dominant theme pick a brief style, then summarizes in technical, business, or policy form.
One model summarizes a source document; a second, independent model checks that every claim is grounded in the source. The clearest demonstration of an independent model as a quality gate against hallucination.
Drives a real local browser: one session searches Google and pulls the first three result URLs, then three parallel sessions each open a URL and capture a screenshot. Tool use with session continuity and fan-out.
Calls a public remote documentation-search tool over HTTP — no subprocess, no API keys. Shows how a remote tool endpoint becomes a single reliable step in the workflow.
A full live-data workflow built entirely from pre-built steps — no custom code. Fetches market data in parallel, parses it, computes the change deterministically, and scores headline sentiment with AI before assembling the analysis.
Answers a question from a local knowledge base, asks the model to cite its sources, then deterministically filters every citation against the documents actually retrieved — hallucinated sources can't survive.
The same retrieval workflow as the previous example, with the retriever swapped to vector similarity over embeddings — the rest of the process is untouched. Shows how the retrieval backend is pluggable without changing the workflow.
A two-stage pipeline where malformed input and a broken business rule are each detected and repaired by a bounded second AI pass — automatically, without failing the run. Deterministic detection, AI recovery.