a Skelf Research project
SQL that learns.
AI that validates.
Privacy that protects.
l0l1 is a developer toolkit for analytics teams who want LLM help on SQL without letting raw rows leave the room. Validation, PII detection, and pattern learning — running where your warehouse already lives.
~/analytics
$ l0l1 check-pii "SELECT ssn, email FROM customers"
Found: SSN, EMAIL_ADDRESS
Anonymized: SELECT <SSN>, <EMAIL> FROM customers
$ l0l1 validate "SELECT * FROM orders WHERE created > NOW() - INTERVAL 7 DAY"
ok schema: orders(id, user_id, total, created)
warn SELECT * pulls 14 columns; only 3 used downstream
$ l0l1 complete "SELECT * FROM orders WHERE"
Suggestions based on 47 learned patterns…01 / VALIDATE
AI as a reviewer, not a writer
l0l1 treats LLM output as a draft. Multi-provider validation (OpenAI, Anthropic) checks queries against your schema before they run.
02 / PROTECT
PII detection on the wire
Presidio-backed detection finds emails, SSNs, phone numbers and credit cards inside queries. Anonymize before sending text to a model.
03 / LEARN
Patterns from your team's queries
Successful queries are stored as patterns per workspace, sanitized of PII first, then surfaced as completions for the next person.
WHAT l0l1 IS NOT
- — Not a text-to-SQL black box. It validates SQL; it doesn't insist on writing it.
- — Not a hosted warehouse. l0l1 connects to yours; data stays put.
- — Not a replacement for your DBA. It's a second pair of (statistical) eyes.
- — Not a magic compliance tool. PII detection is best-effort, not a legal guarantee.
From the blog
All posts → May 22, 2026
What 'privacy' means when your copilot sees your warehouse
Privacy in AI data tools is usually marketing language. Here is a more honest account of what the threat model looks like, what l0l1 actually does, and what it explicitly doesn't claim.
May 1, 2026
Validation as a first-class layer in AI data tools
If LLMs are going to draft SQL, validation has to stop being a UI warning and start being an architectural surface. Here is how we think about it inside l0l1.
Apr 12, 2026
LLM-generated SQL is mostly correct — that's the problem
Why 'looks right and runs' is the most dangerous failure mode for AI-generated SQL, and what a validation layer needs to catch before the dashboard updates.