Data intelligence · Semantic SQL workbench

Core SQL Engineer: Describe What You Need. Get Validated SQL, With Evidence.

Core SQL Engineer is a project-scoped semantic SQL workbench: plain language — or a screenshot — in; validated SQL with safe read-only previews, ER diagrams, schema analysis, staged file imports, and cross-dialect migration out.

Its strongest capability is that it learns your database. Facts you teach it, your business glossary, and your shop's SQL style templates persist per project and shape every future generation — it gets measurably better at your database with use, instead of resetting to zero on every prompt.

And it is safe by construction: every generated statement passes a deterministic parser-based validator — not the model grading itself — and previews run read-only inside transactions that always roll back.

Start a conversation

Natural language in, verified SQL out

Generates ad-hoc queries, views, stored procedures, functions, table DDL, scheduled jobs — and refactors of your existing SQL
Generate mode: a fast one-shot draft built from your schema, glossary, and learned facts
Investigate mode: the AI runs gated read-only probes against the database, verifies its own answer, and returns the SQL with a step-by-step evidence trail
Refine mode: conversational corrections ("wrong schema — use core_plan") that adjust the SQL rather than regenerate it
Live progress streaming — you watch what the engine is doing right now, and can cancel
Every response discloses how many tables the model saw; output is hand-editable and instantly re-runnable

Screenshot → SQL

Attach one or more screenshots of a form, report, or filter dialog. A vision model extracts the structured intent — entities, filters, values — and generates the query against your real schema. Low-confidence extractions open a confirmation dialog before any SQL is generated, and blank fields are never silently turned into filters.

Teach it once, benefit forever

Write facts about your database — where values really live, which join paths are correct — and every future generation uses them
Teach-from-result: after any run, one click turns what was learned into a permanent fact
Thumbs-up confirms a pattern; thumbs-down asks what went wrong and stores the correction
Per-project business glossary maps your organization's vocabulary to the schema
SQL style templates: paste your shop's skeleton once — headers, audit columns, error handling — and every generated artifact reproduces that exact style

Safe by construction

Deterministic validation on Microsoft's T-SQL parser — writes, DDL, and execution are blocked for queries. Not regex, not the model grading itself, not bypassable
Previews are SELECT-only and triple-gated — statement classifier, project policy, connector capability — inside a transaction that always rolls back
The model never sees connection strings or credentials — only a schema summary, your intent, and your glossary and facts
Sharing sample rows with the AI provider is an explicit opt-in, off by default
Row-Level Security support: per-project session context so RLS predicates see the right tenant
Works from a schema snapshot alone — the entire experience runs with no database connection at all, ideal for demos and air-gapped schemas

From hostile files to another platform entirely

File importer — staging first, always

Drag-drop CSV, TSV, or fixed-width; delimiter, encoding, and column types auto-profiled
Fixed-width layouts auto-detected from a sample, then confirmed by you
Complete bundle in run order: staging DDL, load scripts, column mapping, keyed MERGE procedure
Promote to production onto one or more tables with batched MERGE — large loads never long-lock production
Exclude sensitive columns from the entire pipeline
No direct file-to-production merge is ever generated — staging first is a hard rule

Database migration

Six dialects, any pairing: SQL Server, PostgreSQL, MySQL, MariaDB, Oracle, SQLite
Table DDL scripted deterministically from schema metadata — identity, computed columns, defaults, keys, indexes, foreign keys
Views and procedures translated from their captured definitions
Every construct without a direct equivalent gets an inline migration note explaining the difference
Copy per object, copy all, or save the whole migration as one .sql file
Translations are review artifacts for engineers — never executed by the platform

Companion product — CoreSQL Mover. A separately licensed CLI that moves the data itself: incremental SQL Server → PostgreSQL migration built for read-only vendor sources. It detects changed rows without any vendor cooperation, streams them with high-throughput binary transfer, verifies parity table by table, and keeps all of its state on the target — the vendor database only ever receives SELECTs.

See and understand the schema

Mermaid ER diagrams: the whole schema, a named subset, or exactly the tables a generated query touches — rendered next to the results
Deterministic schema-analysis rules: missing primary keys, duplicate and over-wide indexes, foreign keys without supporting indexes, float-used-for-money, and more
Optional workload signals from live servers — missing-index and execution statistics — when a connection permits
Per-project history: search and restore any past run — SQL, preview, diagram, screenshots
Print-friendly export of SQL + preview + diagram

Who it's for

Database engineersData analystsTechnical consultants Vendor-locked teamsHealthcare & insurance data teamsMigration projects

Live schema capture is SQL Server today; the migration workbench translates between all six dialects, and the snapshot connector runs the full experience without touching any database.

Bring your ugliest database.

800 undocumented tables, a vendor that only gives you SELECT, a migration nobody wants to start — that is exactly what it was built for.

Start a conversation