Architecture

How Ward gets from a question to a figure you can check.

Ward queries your systems where they live. Its own data lake runs the math and checks every result. The AI only turns your question into queries and writes the answer, and each figure links back to the row it came from.

One question, end to end

Seven steps, and the model is in two of them.

Most of the work happens in the data lake, where it is cheap and exact. The model is used for the two jobs only a model can do: reading the question and writing the answer.

  1. Connect

    Live databases connect through Ward’s proxy on a read-only credential, and a read-only replica is enough. Exports, Stripe, logs and PDFs sync into the lake by batch, micro-batch or streaming. You choose per source.

  2. Store

    Synced data lands as Parquet-based columnar storage on a custom fork of Apache DataFusion. Columns compress far better than rows, so the same data takes less disk and less memory to scan. Streaming rows are queryable the moment they land.

  3. Index

    Ward keeps its own index format for lookups and query planning, and partitions and clusters the same table at once. A question about one store and one week reads that store and that week.

  4. The model writes the queries

    The model reads your question and the schema it needs, and writes queries in Ward’s JSON query format. It does not see your tables and it does not calculate anything.

  5. The lake computes

    Ward runs the queries, joining live tables through the proxy and synced tables in the lake as one database. Every figure in the answer is computed here, under the asking user’s row- and column-level policies.

  6. Three checks, or it loops back

    Each figure is checked for valid data, an efficient query, and whether it answers the question asked. A failure rewrites the query and checks again. Context windows stay small, so checking adds little to the model bill.

  7. The answer, with its citations

    The model writes the sentence around the computed figures. Each figure is a citation linked to its query and source rows, and every access along the way is logged with the policy that allowed it.

Query in place

Live databases are joined where they sit.

An enterprise will not turn the ocean liner for a few percent. The proxy means it does not have to: point Ward at a read-only replica and ask across every system as one database.

Live databases
Postgreslive databaseno data moved
SAPread-only replicano data moved
Files and SaaS
POS exportfilessynced
StripeAPIsynced
Ward
ProxyJoins live tables where they sit. A read-only replica is enough.
Data lakeColumnar, indexed, partitioned and clustered.
One queryAcross every source, as one database. Each figure linked to its row.

FreshnessSynced sources land by batch, micro-batch or streaming. Streaming rows are queryable the moment they arrive.

The engine

Compression is where the compute bill goes.

Columnar storage, Ward’s own indexes, and partitioning with clustering on the same table. The result is a large lake on a small server.

5 TBof data in one live deployment
5Brecords, growing toward 100B
16 GBserver it is queried on
0migrations to get there

Columnar and compressed

Parquet-based storage on a custom fork of Apache DataFusion. Columns compress far better than rows, so the same data takes less disk and less memory to scan.

Its own index format

Ward keeps its own indexes for lookups and query planning. A question about one store and one week reads that store and that week, not the table.

Partitioned and clustered at once

The same table is partitioned and clustered together, so a query can skip whole partitions and then read a narrow range inside the one it keeps.

The model never does the math

The AI turns your question into queries and writes the answer. The lake computes every figure. The model sees the question, the schema and the result, never your tables, so there is less to pay for and less to make up.

The figures are one live enterprise grocery deployment, not a benchmark. We do not name the operator.

Where the model sits

The model orchestrates. The lake computes and verifies.

JobDone byWhy
Read the questionThe modelLanguage is what models are for.
Write the queriesThe model, in Ward’s JSON formatStructured output the lake can validate before it runs.
Compute every figureWard’s data lakeExact, cheap, and nothing for a model to invent.
Check every figureWard, before the answer returnsValid data, efficient query, answers the question. Failures loop back.
Write the answerThe model, around computed figuresEach figure stays a citation linked to its rows.
Decide who sees whatCedar policy, below the modelNo prompt can widen it. Access control →

You choose the model providers, and a policy decides who may change them. The model receives the question, the schema it needs and query results the asking user may see, never your tables.

Deployment

Your cloud, ours, or on-prem.

Ward’s cloud

Isolated per tenant

Customer data is physically separated per tenant. Only schemas and settings are separated in the application.

Your cloud

Inside your account

Ward runs next to your systems, so the proxy reaches your replicas without leaving your network.

On-prem

Including air-gapped

Runs with no outbound connection, on scheduled updates you bring in.

What is coming

Not shipped yet, and labelled that way.

Setup wizardSchema-change watcherAudit log exportPer-task model routingModel evaluationRegion choice in the SaaSGrafana pluginMaintenance MCPSimulation layer for forecasting
Questions

What reviewers ask about the architecture.

Not your live databases. They connect through Ward's proxy and are joined in place, so a read-only replica is enough and nothing is moved. Sources that are not databases, like exports, Stripe, logs and PDFs, sync into Ward's data lake. You choose per source.

Parquet-based columnar storage on a custom fork of Apache DataFusion, with Ward's own index format for lookups and query planning, and partitioning and clustering on the same table.

Live databases are as fresh as the replica Ward reads. Synced sources land by batch, micro-batch or streaming, and streaming rows are queryable the moment they arrive.

One live enterprise grocery deployment queries 5 TB and 5B records on a 16 GB server, and is heading toward 100B records. That is a deployment, not a benchmark.

Your cloud, Ward's, or on-prem, including air-gapped deployments with scheduled updates. In Ward's cloud, customer data is physically isolated per tenant.

You do. Ward offers starter schemas for common sources if you want a head start, and nothing about your schema is locked to Ward.

Send this to whoever will review it.

Then connect one read-only source and check the first figure yourself.

Read-only to start · your LLM keys · SOC 2 Type II underway · or book a call directly

Find out what your data has been hiding.

Tell us about your operation. We’ll show you the problems Ward catches, and the ones your current tools miss.

Step 1 of 3
What are your goals?
Step 2 of 3
About your operation
Step 3 of 3
Your contact info