Engineering Blog
Design decisions and mechanisms inside the engines, and the services opened on them. Dates and numbers come from repository history and recorded source inspections.
Latest
The model proposes, the system decides: the design and operating method four engines share
Text-to-SQL, controlled generation, knowledge publishing, document reasoning: why four engines in different fields stand on the same three rules, and how those rules are kept in code and operations.
Before a question becomes SQL: a Text-to-SQL engine built on semantic contracts and AST validation, and the QueryMind deployments it runs
The hard part of turning language into SQL is not generation but meaning and boundaries. Injecting metric contracts into the syntax tree, rejecting forbidden nodes before execution, running read-only — the design and five months in the field.
Publishing approved queries as MCP tools: QueryMind Connect, which never regenerates SQL per call
The most dangerous way to open enterprise data to an AI client is to let the model write SQL on every call. Here only administrator-frozen definitions become tools, every call runs under argument, permission and output limits — and a real call recorded on November 8, 2026.
The database decides, the model composes: a controlled-generation engine built on four packet layers, and the Taste Edition service opened on it
In a long-form generation system some decisions must not be left to the model: what to reference, in which order and length to write, whether a draft passed. How D1–D4 distillation packets, database gates, checks and a publish gate moved those decisions into the system — and forty days of unattended serialization.
From one approved snapshot to HTML, JSON-LD, facts.json and answer pages: a structured-knowledge publishing engine, and the FactMind service opened on it
The facts AI search reads must come from the same approved content people read. A site-control gate, five kinds of output from one snapshot, consistent corrections, and a loop that measures exposure in numbers.
A rule engine that keeps 'unknown' as a result state: document reasoning for court auction files, and the Jupgo service opened on it
Confirmed absence and unobtained information are different inputs. Extracting tenant lists as three states — absent, unknown, value — a decision function that calls no model, conservative assumptions written into the result, and the record from app to MCP.
Why SQL validation starts before SQL.
Meaning, scope, and permission are separate concerns.
Why approval is an execution boundary.
A prepared query still needs current access.
What makes a data answer inspectable?
Result grounding and the limits of numeric checks.
Technology milestones
Only dates confirmed in repository history.
- MCP publishing
Approved-query MCP call recorded — QueryMind Connect in Claude
- Text-to-SQL
First real-customer database freeze reported
- Knowledge publishing
GEO engine deployed — AI answers, sources, exposure engine, answer pages, crawl queue
- MCP publishing
Controlled-generation MCP stage one — six Taste Edition read tools
- MCP publishing
Document-reasoning MCP connector — Jupgo in Claude and ChatGPT
- Controlled generation
Unattended serialization, one episode a day — Taste Edition episode 18
- Controlled generation
Four database gates added to the controlled-generation engine
- Knowledge publishing
Knowledge-publishing engine public — FactMind homepage; first self-service registration the next day
- Text-to-SQL
Oracle-based university records database connected
- Text-to-SQL
First field deployment of the Text-to-SQL engine (QueryMind)
- Text-to-SQL
Metric contracts injected by the server into the SQL syntax tree (request → contract)
- Document reasoning
Document-reasoning engine first commit — auction prediction and serving
- Engineering
On-premise small-LLM provider (OpenAI-compatible) supported
- Text-to-SQL
Text-to-SQL engine repository begins