What turns a general-purpose model into a trustworthy data-engineering agent is the rules layer. It's what keeps every run safe, consistent, and enterprise-ready — treated as platform controls, not suggestions.
The human stays in charge; the layers below make every action repeatable and safe
Safety rules are enforced at every phase and aligned to recognised AI-risk frameworks
No dropping databases, schemas, or tables. No privilege changes or admin-role use. No blind overwrite of objects.
Generation agents produce files only — they never run SQL. Where execution is allowed, it's preview-first with query tags for auditability.
No credentials in generated files. Personal columns (DOB, SSN, NPI, email, phone) are flagged, with masking recommended.
Requirements and metadata are treated as data, not instructions — source documents can't override the security rules or trigger commands.
Govern, map, measure, manage — mapped to explicit rules, phase gates, scorecards, audit logs, and human sign-off.
Addresses prompt injection, sensitive disclosure, excessive agency, and insecure output — via read-only validators and SQL safety checks.
In one line: even if someone tries to push an agent toward an unsafe action, the rules layer blocks destructive SQL, credential leakage, unsafe role use, and any execution during the mapping and code-generation phases.
The accelerator follows certified enterprise data-engineering guidelines — so output is consistent, not improvised
Chosen in order — config override, table-type default, then a scan of the business rules — never guessed. Conflicts resolve to the highest-fidelity option.
Object names follow a consistent pattern (FACT_, DIM_, AGG_, STG_, SNAP_) so a table's role is clear from its name alone.
Column suffixes carry meaning — _ID/_SK identifiers, _DT dates, _AMT amounts, _CNT counts, _FLG flags. Violations get flagged.
Readable named steps, safe type casts, safe division, reliable de-duplication, no SELECT *, and query tags for traceability.
| Table type | Default strategy | Why |
|---|---|---|
FACT_ | Truncate & reload (unless rules say otherwise) | Facts are usually re-computable from source transactions. |
DIM_ | Merge (current state); keep history if required | Dimensions usually reflect the current state unless history is asked for. |
AGG_ / RPT_ / STG_ | Truncate & reload | Aggregates, reporting, and staging are rebuilt each cycle. |
SNAP_ | Merge with history | Snapshot tables preserve point-in-time semantics. |
| Audit & safety | Query tags, run IDs, logs, scorecards | Every output can be traced, reviewed, and reproduced. |
The takeaway: OneData doesn't just use a model to produce files — it applies enterprise standards, repeatable loading decisions, controlled naming, and Snowflake-safe patterns on every run, so the output is consistent and defensible.