> ask your warehouse anything

Natural-language access
to data warehouses. In private beta.

Production-grade NL→SQL for Snowflake. Schema-aware. Audit-clean. Cost-conscious.

Get a demo → Read the architecture ↓
BUILT WITH · Snowflake · dbt · Airflow · LangGraph
Snowflake-nativeSQL + metadata aware
dbt-manifest awareModel lineage in planner
LangGraph orchestrationMulti-step query planning
4-layer plannerSchema-grounded reasoning
Audit by defaultEvery query logged
STACK WE SHIP IN PRODUCTION
Snowflake ― dbt ― Fivetran ― Airflow ― LangGraph ― AWS ― Python ― Snowflake ― dbt ― Fivetran ― Airflow ― LangGraph ― AWS ― Python ―

Real tools. Real recognition. Honest framing — we ship in these stacks. No partner claims.

01.

Natural language,schema-grounded.

The planner is dbt-manifest-aware, joins-aware, semantic-layer-grounded. Before the model writes a single line of SQL, it narrows the question to the tables an engineer who knew the warehouse would have chosen. Produces SQL a human reviewer would sign off on.

Query result — chart, table, generated SQL
02.

The customer tableproblem.

Real warehouses have customers_v2_final_USE_THIS next to customers_v2_final. Our planner figures out which is canonical before the model sees the question.

The signal
dbt manifest + INFORMATION_SCHEMA + query history, scored as graph weights.
Schema view — canonical table highlighted, deprecated forks dimmed
03.

Know the costbefore you run it.

Preview cost before execution. Stops the runaway dashboard-spam problem.

SNOWFLAKE_S
Scan Volume
120 MB
Est. Cost
$0.02
SNOWFLAKE_M
Scan Volume
640 MB
Est. Cost
$0.08
SNOWFLAKE_XL
Scan Volume
2.1 TB
Est. Cost
$0.42
Cost preview modal — 2.1 TB scan, $0.42 estimate, proceed or rewrite

FIG. 03 · COST PREVIEW MODAL — BEFORE EXECUTE

04.

Every question logged.
Every query traceable.
No shadow analytics.

TimeUserQuestionCostWH
2026-05-12analyst_3"gross margin by category"$0.04M
2026-05-12ops_lead"events_2024 row count"$0.17S
2026-05-12cfo_review"q4 churn rate"$1.23XL
2026-05-12analyst_3"active users last 7 days"$0.06S
2026-05-12ops_lead"p99 latency last 24h"$0.02S
→ See the full audit log mockup ↓
Audit view — query log with user, SQL, cost, timestamp
05.

One semantic layer.Different SQL dialects.

Snowflake is live in private beta. Databricks is in active build. BigQuery and Redshift on the roadmap. Same semantic layer, every time.

Multi-warehouse switcher — Snowflake live, Databricks beta, BigQuery roadmap
FIG. 06 · ARCHITECTURE

How the planner thinks.

Four layers. Each one is something an engineer who has lived inside a real warehouse would recognize.

Architecture diagram — 4-layer flow from NL interface through planner, semantic layer, to warehouse executors
FIG. 07 · ENGINEERING NOTES

How we survive real warehouses.The schema introspection problem.

Every NL→SQL demo I have ever watched gets the easy half right. It picks the obvious table, joins on the obvious key, and answers a question about an orders table that has eight rows and three columns. The hard half lives a layer deeper — six orders tables, four deprecated, two disagreeing about what "completed" means.

Read the full engineering note →
FIG. 08 · ABOUT THE TEAM

HKDataCrest is a data and AI engineering team based in India, building for US companies. Snowflake, dbt, Fivetran, Airflow, and LangGraph in production. We work on scoped builds and embedded engineering for ops, finance, and data leaders at mid-market companies. This natural-language warehouse layer is the worked example of how we think about the data layer underneath AI products.

Snowflake dbt LangGraph Airflow Fivetran AWS Python GST-Registered MSME-Certified
Start with a Free Audit Email Arshit
PRIVATE BETA · LIMITED INTAKE

Stop hand-rolling SQL.
Start with a Free Audit.

A 5-page memo on your data stack — what to keep, what to retire, where the next AI surface lives. No sales call required.

Book a discovery call → Request audit by email
5-DAY TURNAROUND· NO OBLIGATION· SALES@HKDATACREST.COM