Skip to content
Self-improving NL2SQL engineby Irevex

Ask your database anything.

DataSpeak turns a question into SQL against your own database, shows you exactly what it ran, and keeps every correction your team approves as grounding for the next question.

No credit card required Free 14-day trial Cancel anytime

Someone in ops asks

“What was revenue by region last quarter?”

Production DB

DataSpeak first wrote SUM(o.total) as revenue. An analyst flagged that this counts refunded orders, and corrected it to SUM(o.total minus o.refunded). The approved correction was written to the dataspeak_knowledge collection, so later revenue questions on this schema start with it applied.

Illustrative schema and figures. The correction cycle is real.

99.9%

Uptime

<2s

Avg response

50K+

Queries processed

500+

Active teams

Unique capabilities

What makes DataSpeak different.

Other tools that write SQL start from zero on every question. This one retrieves what your people have already settled, then writes the query.

Self-improving AI

Every query and every judgment feeds a vector knowledge base, so accuracy climbs with use instead of staying flat.

Enterprise security

Fernet encryption for API keys and database credentials. Nothing sensitive is stored as plain text.

Universal database support

PostgreSQL, MySQL and SQLite, with automatic schema discovery on connect.

Real-time feedback loop

Users rate results. Correct answers strengthen the pattern; wrong ones open a review workflow.

Vector knowledge store

Qdrant-backed semantic search over query patterns, table instructions and domain knowledge.

Granular permissions

Five role tiers from Owner to User, assigned per instance so access follows the database.

Continuous learning

The self-improving feedback loop.

Every query and every judgment makes the next answer better. A single thumbs-down never edits the model on its own — a person signs off first.

Natural language query

Someone asks a question in plain English.

AI SQL generation

A LangGraph pipeline writes SQL against the discovered schema.

Execute and display

Results come back with the SQL, the table and the answer.

User feedback

Thumbs up or down, with an optional reason.

Vector store learning

The judgment is embedded into Qdrant for later retrieval.

Continuous improvement

Approved corrections ground the next similar question.

Enterprise security

Your data is encrypted and protected.

Connection strings, database passwords and LLM keys are encrypted at rest with Fernet — AES-128 in CBC mode — never stored as plain text. Sessions run on JWTs with refresh tokens.

API key encryption

Fernet symmetric encryption for every LLM API key.

Password encryption

Database credentials encrypted at rest.

JWT authentication

Token-based auth with refresh tokens.

Role-based access

Instance-level permission control.

OwnerAdminEditorAnalystUser
LLM API key
sk-••••••••••••••••••••••••
Encrypted
Database password
••••••••••••••••••••••••••
Encrypted
Connection URL
postgresql://••••@••••:5432/••••
Encrypted

All sensitive values encrypted with Fernet (AES-128, CBC).

Encryption specs

FernetAES-128CBC modeJWT + refresh
Rich results

See results exactly how you need them.

A plain-English answer, the SQL that produced it, and the rows themselves — side by side, so the analyst and the operator can both trust it.

Show me the top 5 customers by revenue this month

DataSpeak

Here are the top customers by revenue for this month.

01SELECT c.name, SUM(o.total) AS revenue
02FROM customers c
03JOIN orders o ON c.id = o.customer_id
04GROUP BY c.name ORDER BY revenue DESC;
namerevenue
Acme Corp$125,430
TechStart Inc$98,210
Global Systems$87,540
Synthetic example — 3 of 5 rows · customers, orders

An answer in words

Written for whoever asked, not for whoever maintains the schema.

The SQL that ran

Shown by default, never hidden — so it can be checked and corrected.

The rows themselves

Formatted tables with the tables touched and the time taken.

What ran, in numbers

Row count, the tables touched, and how long it took — so a result is never a black box.

Copy the SQL

One click hands the exact query to anyone who wants to run it, tune it, or drop it straight into a report.

Analytics

Find out where your schema needs a note.

Accuracy is broken down per table, so a column people keep misreading shows up as a number rather than as a complaint. Write one table instruction, and every later question about it starts from that.

Accuracy tracked per table

See which tables the AI already answers well, not just one aggregate score.

Flagged where accuracy slips

A table that answers worse than the rest is usually one nobody has documented.

Response times per instance

Find the connection that has started to drag before anyone reports it.

1,247
Questions asked
94.2%
Answered correctly
2
Tables needing a note

Accuracy by table

orders97%
subscriptions78%
refunds71%

subscriptions and refunds answer worse than the rest. Both were corrected for the same reason — add a table instruction and the pattern stops recurring.

Synthetic example data
Multi-database management

Manage all your databases in one place.

Connect PostgreSQL, MySQL and SQLite instances, each with its own access control and connection monitoring.

Production DB

PostgreSQL

LiveOwner

Analytics Warehouse

MySQL

LiveAdmin

Test Database

SQLite

OfflineEditor
Synthetic example instances

Role-based access

Owner, Admin, Editor, Analyst, User.

Real-time status

Live connection monitoring with health checks.

Schema viewer

Browse tables, columns and relationships.

Vector knowledge base

Powered by Qdrant vector search.

DataSpeak does not just generate SQL — it remembers. Every successful query, every correction and every judgment is embedded and retrievable.

Semantic retrieval

Find similar past queries using embeddings.

Table instructions

Custom notes per table for better queries.

Approved corrections

Learn from reviewed negative feedback.

Gap analysis

Compare a query against what is already known.

Vector collections

Qdrant
dataspeak_knowledge

Successful queries and their patterns

2,847 vectors
dataspeak_instructions

Table-specific instructions

156 vectors
dataspeak_feedback

Every judgment, embedded

423 vectors

Embeddings: Google gemini-embedding-001 via LangChain.

Vector counts are illustrative
Feedback & review

Collaborative query improvement.

Users give instant feedback. Maintainers review and approve corrections. Nothing reaches the model without a human signing off.

Positive feedback

Strengthens patterns that already work.

Stored in the vector knowledge base

Retrieved for similar questions

Improves future SQL generation

Negative feedback

Opens the review workflow.

Requires maintainer review

Approved corrections guide the AI

Prevents the same mistake twice

Review workflow

01Pending02Under review03Approved04Applied to AI
Everything you need

Powerful features that make your data accessible to everyone.

Natural language

Ask in plain English

SQL generation

See exactly what ran

Results display

Tables with the SQL beside them

Self-learning

Improves with every query

Multi-database

Postgres, MySQL, SQLite

Secure

Encrypted credentials, JWT auth

Team collaboration

Granular role-based access

Analytics

Usage insights and metrics

Fast

Answers while you wait

Vector store

Semantic knowledge base

Feedback loop

Continuous improvement

Instance management

Many databases, one place

Ten minutes to test

Ready to transform how your team accesses data?

Connect a database and ask it something badly. The first answer might be wrong. Correct it, and watch the second one arrive already knowing.