Autonomous Intelligent System
Ask a codebase a question. Get an answer with receipts.
AIS reads an unfamiliar project, builds a model of how it is put together, and answers architecture questions — with a file and line reference behind every claim it makes.
A live instance is running at ais-demo.spru.run. It runs the real engine against a public third-party codebase (TypeORM core, MIT): discovery, module graph and file:line evidence. Prose answers need an LLM key, so this instance answers from the deterministic half of the pipeline and labels itself accordingly.
> Where does the answer get checked? Each answer passes through the evidence loop before it is returned. The loop keeps the claim, the file it came from, and the lines that support it. src/core/evidence-loop/evidence-loop-service.ts 61 export class EvidenceLoopService
Four steps, no guessing
AIS is not a chat window over your repo. It walks a pipeline, and every stage leaves a trace you can inspect.
Pose the question
Plain language, about how the project is built rather than what a file contains.
Map the structure
The discovery pipeline walks the tree and builds an architecture model of the project.
Answer with claims
The engine returns a structured answer decomposed into separate, checkable claims.
Show the receipts
Every claim is tied back to source lines, so you can confirm or reject it yourself.
What an answer looks like
Real examples from a public open-source codebase. The grey strips are the evidence: file, line, and the source it was drawn from.
How is a user question routed through the analyzer?
From the outside it looks like one call, but the orchestration layer separates the request from the work that answers it.
The orchestration layer owns the conversation shape. It hands the question to the engine as a task and hands the result back as a response, so the engine never talks to the user directly.
Where is confidence decided?
Not in a single score, but in a set of contracts that the pipeline and the engine both agree on.
Confidence lives in a contract module, separate from both the pipeline that produces the result and the engine that consumes it. That keeps the threshold policy in one place instead of spread across call sites.
Can AIS run without a model key?
Yes — it starts and serves the interface, but real answering needs a provider configured.
Without AIS_EXECUTION_REAL=true, AIS_REAL_LLM=true and OPENAI_API_KEY, the UI starts but the execution engine stays in its non-real path and will not produce genuine answers.
Where it still hurts
AIS is a working prototype, not a finished product. The failures are documented as carefully as the wins.
- Retrieval is keyword-based, not semantic — unusual phrasing can miss the right file.
- Answers can be wrong. Treat every claim as a hypothesis to check, not a conclusion.
- Strongest on TypeScript and JavaScript projects.
- Analyses local directories only.
- Large monorepos over ~10k files discover slowly.
Run it on your machine
The engine is MIT licensed. The source is kept private and shared on request — as a review archive, for evaluation or for a paid engagement.
Write to ask for access and describe the project you want analysed; archives go out per request.
# after you receive the source archive unzip ais-workspace.zip && cd ais-workspace npm install && npm run build # start the AIS UI on port 3456 npm run mvp-ui # optional: enable real inference export AIS_EXECUTION_REAL=true export AIS_REAL_LLM=true export OPENAI_API_KEY=your_key
SIP ships as Sec Scanner
SIP — the Security Intelligence Platform that came first — is now an auxiliary product inside AIS, and it is the commercial part of the project. AIS itself stays free and MIT licensed; the source is shared on request.