About l0l1
l0l1 is a project from Skelf Research. It exists because we kept watching the same scene play out: an analytics engineer asks ChatGPT to fix a join, the model returns confident SQL, everyone says “ship it,” and the query silently returns wrong numbers for a week.
The position
Large language models are good at generating SQL that compiles. They are not good at generating SQL that answers your actual question against your actual schema with your actual semantics. That gap is the whole problem.
l0l1 is built on a simple premise: the model writes a draft, and a
deterministic layer reviews it. Schema-aware validation,
sanity checks for SELECT *, ambiguous joins, and missing
WHERE clauses, and PII detection at the prompt boundary so
sensitive literals don't leak into provider logs.
What it actually does
- Validate SQL with AI providers (OpenAI or Anthropic) using schema context you pass in.
- Detect PII in queries using Microsoft Presidio plus extra regex patterns for SSN, email, phone, and credit card.
- Anonymize sensitive literals into placeholders before queries are stored or sent upstream.
- Learn patterns from successful queries per workspace, then suggest completions based on those patterns.
- Connect to PostgreSQL, MySQL, SQLite, and DuckDB.
The full surface area is in the documentation.
How to use it
There are four front doors: a Typer-based CLI, a FastAPI REST server,
a Jupyter magic, and a VS Code / LSP integration. Pick the one your team
is already in. Most people start with l0l1 validate from a
shell and graduate to the API once they want it inside an existing data
platform.
What we won't claim
We don't ship differential privacy. We don't promise zero false negatives on PII detection — nobody can. We don't replace your data governance program. l0l1 is a careful, opinionated layer between an LLM and a warehouse, not a compliance product.
License and source
MIT licensed. Source at github.com/skelf-research/l0l1-api. Issues, PRs, and uncomfortable questions are all welcome.