Kai · Coming soon

Meet Kai.
A curious mind for your data.

Some databases are hard to understand. We’re building Kai to explore how your data fits together, ask when something is unclear, and show where each answer comes from.

Available today: Akuma, Enzan and Sōzō through Kaizen Platform.

Kai concept preview

Here is how Kai is designed to work: explore the structure first, show what it cannot resolve, confirm the meaning with you, then explain the answer with the query and records behind each figure. The frame below is a concept with invented data.

KaiComing soon

Illustrative data

Your business question

Which plans drove revenue last month?

  1. 1 · Explore the schema

    Kai is designed to read the structure first and propose what each table and column holds, marking what it cannot resolve on its own.

    • inv_hdrHypothesis

      One row per issued invoice

    • pay_txnHypothesis

      Payments received against invoices

    • plan_refHypothesis

      The plan names the figures are grouped by

    • acct.plan_cdUnresolved

      Looks like the link to plan_ref.cd, but no foreign key declares it

    • inv_hdr.statUnresolved

      Holds 1, 2 and 9; 9 may mean voided

    Fictional schema. Every name and meaning here is invented for the design; these are hypotheses Kai would propose, not verified facts.

  2. 2 · Identify what is uncertain

    Before any figure is computed, these would come back to you:

    • Is acct.plan_cd the join to plan_ref.cd?
    • Does stat = 9 mean a voided invoice?
    • Revenue can mean invoices issued or payments received. Which definition fits your question?
  3. 3 · Confirm the meaning

    Preconfirmed for this illustration

    • Confirmedacct.plan_cd joins plan_ref.cd.
    • Confirmedstat = 9 is a voided invoice, so it is excluded.

    Answers written for the design to the two questions in step 2; in the planned workflow, you would give them.

    Your choice: revenue definition

    Changing the definition changes the figure, the change rate, the explanation, the source line and the table in step 4. It does not re-check the two facts above.

    Invoiced revenue: $386,000 · Cash collected: $350,000 for August, illustrative.

  4. 4 · Explain the answer

    Because you chose Invoiced revenue · August

    $386,000↑ 10.3% from July

    Scale and Growth added $45,000, offset by a $9,000 decline in Starter.

    Invoiced revenue would sum inv_hdr by plan through the confirmed plan_cd join, with voided invoices excluded.

    Illustrative invoiced revenue by plan, July and August
    PlanJulyAugustChange
    Scale$194k$216k+$22k
    Growth$105k$128k+$23k
    Starter$51k$42k−$9k

    Example billing data

    Inspect the reasoning
    Kai is designed to show the tables, joins and filters behind each figure once the meaning is confirmed. Every value here was invented for the design; no query was run and no schema was read.

    Planned

    The planned workflow keeps confirmed meanings with the analysis so later questions can build on them; changed definitions or schema require review.

Proposed interface with illustrative data. Not live product output.

Available today

Start building with Kaizen Platform today.

Akuma answers questions today about a database you connect, while Kai is being built for databases nobody has explained yet, to work out their meaning with you first. Use the dashboard, the API and SDKs, or MCP for selected tools, with one account and API key.

  • Akuma

    Ask a connected database. Inspect the SQL and results.

  • Enzan

    Compare current GPU offers. Understand recorded Kaizen AI costs.

  • Sōzō

    Generate sample data for tests and demos.

Create account

Access Akuma, Enzan and Sōzō through Kaizen Platform.

Read the docsTalk to sales

After signup and sign-in, the dashboard opens in the Akuma workspace; an email signup confirms through the emailed link first. From there: connect an approved database, open Live Pricing, or generate a preset dataset. Creating an account does not include Kai.

01 / Akuma

Your dashboard doesn’t answer today’s exception question.

Ask your database. Inspect the SQL and results.

For reporting and operations owners whose exception question changed after the dashboard was built, and the engineer who can approve a read-only source for them.

  1. Ask a scoped question about a connected PostgreSQL or MySQL source.
  2. In the interactive query workflow, Akuma can ask which definition you mean when it finds a question ambiguous.
  3. Inspect the read-only SQL, its rows, and the tables and filters it used.

AkumaKaizen Platform · Available today

Illustrative data

Your question

Which paid orders have waited more than 3 days without fulfilment, excluding cancelled and refunded ones?

PostgreSQLRead-onlySQL + results

4 orders need attentionas of 13 Sep 2026

This question named its own definitions, so it ran straight through. In the interactive query workflow, a question Akuma finds ambiguous may come back with options to choose from instead of SQL.

Next step: review these orders in your own fulfilment process. Akuma reads the database; it does not change orders or contact customers.

Illustrative paid orders waiting more than 3 days without fulfilment
OrderPaid onDays waitingRegionStatus
ORD-104822026-09-058EUprocessing
ORD-104972026-09-076USpending
ORD-105112026-09-085APACprocessing
ORD-105302026-09-094USpending

Example order database · PostgreSQL · fictional names

Inspect the SQL and provenance

Written for this page to match the table; no Akuma call produced it.

SELECT o.order_number, o.paid_at::date AS paid_on,
       DATE '2026-09-13' - o.paid_at::date AS days_waiting,
       o.region, o.status
FROM orders o
WHERE o.paid_at IS NOT NULL AND o.fulfilled_at IS NULL
  AND o.status NOT IN ('cancelled', 'refunded')
  AND DATE '2026-09-13' - o.paid_at::date > 3
ORDER BY o.paid_at;

Provenance comes from the executed SQL and the engine's own table list, never from the model's explanation. When present on a completed answer, provenance is labeled full, with tables and filters parsed from the SQL, or limited, with available table metadata and a stated reason.

Illustrative interface and data, not a screenshot. Akuma is available today through Kaizen Platform.

Prerequisites and next step

Live results need a connected PostgreSQL or MySQL source. No connection yet? Use an approved schema to draft SQL and run it in your own tools.

  • Who does what. The reporting owner brings the question and its definition; an engineer or administrator supplies the approved source.
  • Access. Connect with a database role that has only the read permission you intend. Dashboard queries run read-only, and API requests can scope tables, columns, row counts and timeouts.
  • Connectivity and eligibility. Live sync supports PostgreSQL and MySQL. The connector rejects private-network and loopback hosts, and a source whose schema exceeds the eligibility limits is reported as too large rather than connected. These are the main checks, not a complete setup guide; confirm reachability with your engineer.
Create account

After you create an account: in the Akuma workspace’s Data Sources view, connect an approved PostgreSQL or MySQL source or save a manual schema, then ask a first question.

02 / Enzan

GPU quotes use different configurations and pricing units.

Know what a GPU quote really buys.

For the ML engineer or infrastructure buyer comparing quotes that describe different hardware, instance sizes and units. A shortlist here is research, not a promise of the cheapest or available capacity.

  1. In Live Pricing, filter current offers by GPU type, region, minimum number of instances, training readiness and an hourly ceiling.
  2. Read each rate per buyable instance, with the GPUs it includes, its source and its freshness.
  3. Checking a quote? Through the API and SDKs, the same search can also rank offers against the hourly rate you were quoted.
  4. Verify terms and capacity with the provider before you commit.

EnzanKaizen Platform · Available today

Illustrative data

Your requirement

Which offers fit this training job, and what does it cost per hour?

H100 · 8 GPUs per instanceRequested: 2 instancesTraining-readyCeiling $26.00 per instance-hour

Compare offers in the same units.

Per buyable instance-hour, cheapest first, with each source's freshness beside it.

Fictional offers matching the displayed requirement, cheapest first, with each offer’s hourly total for the requested instances
Example offerGPUs per instanceFreshnessPer instance-hourPer hour · 2 instances
Provider A8fresh$20.80$41.60
Provider B8fresh$23.20$46.40
Provider C8stale$25.60$51.20

16 GPUs in total: 8 per instance across 2 instances. The $26.00 ceiling filters each instance's rate; for 2 instances that is at most $52.00 per hour. Hourly totals are arithmetic from this fictional example, not a job budget: duration and other provider charges are separate. Live Pricing shows per-instance rates with each offer's source and freshness.

Fictional offers · not live prices

Understand the totals
Provider A's $20.80 per instance-hour is $2.60 per GPU-hour, and 2 instances are $41.60 per hour. Stale means the source has not refreshed recently. Nothing here is a live price, inventory or reserved capacity.

Illustrative interface and data, not a screenshot. Enzan is available today through Kaizen Platform.

Prerequisites and next step

Live Pricing is in the dashboard after you create an account. Ranking against a quoted rate is an API and SDK search option.

Create account

After you create an account: open the Live Pricing tab of the Enzan dashboard and set your filters.

AI cost analytics

See which models drive your AI costs.

Understand your usage. Choose your next step.

Enzan breaks down estimated costs for AI activity recorded through Kaizen. Evaluate recommended changes, configure alerts, and match Akuma’s model to question complexity.

EnzanKaizen Platform · Available today

Illustrative data

AI cost analytics · recorded activity · last 7 days

large-model is 65% of the estimated cost from 18% of queries.

Estimates for activity recorded through Kaizen, by model and window. Illustrative figures. Not imported provider invoices, and not usage outside Kaizen.

Illustrative estimated cost by model over the last 7 days, with the total
ModelQueriesEstimated costShare
large-model1,300$39.0065%
mid-model3,900$16.8028%
small-model2,000$4.207%
Total7,200$60.00100%
Cost by model API docs

Illustrative interface and data, not a screenshot. Enzan is available today through Kaizen Platform.

Cost optimization recommendations

Find the next change worth testing, with its conditions stated.

Each recommendation says what Enzan found and proposes one change, with a heuristic confidence score and estimated monthly savings. Nothing is applied for you.

Illustrative shape from the API contract; no recommendation here was produced for a real account.

title
The change found, such as moving one model's simple prompts to a cheaper model.
description
What Enzan found in your recorded activity: the share of that model's queries that are simple prompts.
estimatedSavings
The modeled monthly impact at catalog rates.
confidence
A heuristic score from 0 to 1 set by the rule, not a measure of answer quality.
suggestion
The proposed change: route those simple prompts to a cheaper model from the same provider.

Your evaluation: test answer quality on representative questions before changing routing. Enzan does not run that test for you.

Enzan reviews your recorded activity for a change worth testing. A model-downgrade recommendation appears only when that activity shows all three:

  1. Simple prompts are a meaningful share of one model's queries.
  2. The same provider has a cheaper model for that prompt mix.
  3. Moving those prompts would cost less at catalog rates.

The cost table shows cost by model only, so it does not show whether a downgrade applies.

Recommendations API docs

Cost and pricing alerts

Get a webhook when a configured condition needs attention.

Needs a configured rule and webhook destination, managed in the dashboard's Alerts tab or through the API. Cost rules use recorded account activity; pricing rules track Kaizen's LLM/GPU rate catalogs, not live offers. Alerts fire only where alert evaluation is enabled. An alert notifies; it does not stop spend or change your configuration.

  1. 1 · Rule One of the six creatable kinds below, with that kind's settings. Webhook endpoints are configured separately; every enabled endpoint in your account receives every alert event, whichever rule fired it.
  2. 2 · Qualifying event Recorded when alert evaluation runs, the rule's condition holds and the event is new; it carries the rule id, type and time. A persisting condition does not record an event on every run: a cost threshold rule records nothing new until its window has passed since its last recorded event, and a duplicate of an already recorded event is not recorded again.
  3. 3 · Webhook delivery One delivery is queued per enabled endpoint and may be retried; it is reported as pending, sent or failed, with a retry count.

Illustrative sequence, not a recorded event or delivery.

Cost thresholdBudget exceededCost anomalyOptimization availablePricing changeDaily summary
Alerts API docs

Akuma model routing · API/SDK

Match the model to the question's complexity.

Route simple, moderate and complex Akuma questions to different configured models, then see the estimated difference against the default model. Configured through the Enzan API and applied only once you enable it; evaluate answer quality on your workload.

Illustrative policy, not a default: the fields one routing update sends, not the configuration the API returns. Nothing here was sent or enabled.

enabled: true
simple_model: small-model
moderate_model: mid-model
complex_model: large-model
Routing API docs

Prerequisites and next step

Uses activity recorded in your account for the selected window. Recommendations estimate opportunities; you decide which change to evaluate.

Sign in

Already using Kaizen? Sign in: the Enzan dashboard shows recorded activity by window, the Alerts tab manages rules and webhook endpoints, and routing is configured through the API and applies only once you enable it.

03 / Sōzō

You need useful demo data without copying customer records.

The right data to build with.

For a developer or solutions engineer preparing staging, a test or a demo without copying customer rows: generate records shaped for the job, inspect them, and load the fixture yourself.

  1. Start from a preset: SaaS customers, e-commerce orders or users.
  2. Or define fields through the API: UUIDs, choices, integer and float ranges, booleans and relative dates.
  3. Inspect rows and per-column statistics, then take CSV or JSON Lines into your fixtures with the SDK helpers.

SōzōKaizen Platform · Available today

Illustrative data

Your request

Create customer data for our staging app.

1,000 recordsseed 42CSV via the SDK helper

Defined schema

customer_id: uuid4
plan: choice:starter,growth,scale
region: choice:us,eu
seats: int:1-200
active: boolean:0.8

Sample rows that fit.

Define the shape your application expects. Generate synthetic records for development, testing and staging.

Synthetic sample rows for the example customer schema
customer_idplanregionseatsactive
3f9c…a1growthus42true
8d21…7estartereu6true
c04b…59scaleus180false

Synthetic sample · no customer records

About this sample
Illustrative rows inside the schema's choices and ranges, not a generated download. IDs shortened.

From response to fixture

  1. 1 · Generate A dashboard preset, or a schema you define through the API.
  2. 2 · Inspect Rows and per-column statistics in the response.
  3. 3 · Convert The API response is JSON; the SDK helpers turn it into CSV or JSON Lines for your fixture files.
  4. 4 · Load You import the file into your application or test tooling. Sōzō does not write to your systems.

Dashboard presets

Built-in schemas to start from in the dashboard. Through the API, a schema you define works in their place.

SaaS CustomersE-commerce OrdersUsers

Illustrative interface and data, not a screenshot. Sōzō is available today through Kaizen Platform.

Prerequisites and next step

Presets in the dashboard; custom fields through the API.

Create account

After you create an account: choose a preset in the Sōzō workspace and generate; define custom fields through the API.

For developers

Build with
Kaizen Platform.

From a question to an API call. Bring Akuma, Enzan, and Sōzō into the tools your team already uses.

  • One API key across the platform
  • TypeScript, Python, and Go SDKs
  • MCP server for agent workflows

SDK and MCP source on GitHub. The examples use these clients from source.

MCP tools include akuma.query, enzan.summary, sozo.generate, sozo.schemas.

Embedding database Q&A in your app

The interactive query endpoint can return completed, needs_clarification or rejected, so your app can ask the user before SQL runs.

API examples

akuma-query.ts

SDK sourcegithub.com/kaizen-ai-systems/sdk-typescript

Job
Ask a scoped revenue question of a connected PostgreSQL source, read-only and limited to the invoices and plans tables.
What this prints

Prints the generated SQL, then the returned rows as a console table; an empty table when the response has no rows.

import { createClient } from "@kaizen/sdk";

const kaizen = createClient({ apiKey: process.env.KAIZEN_API_KEY! });

const result = await kaizen.akuma.query({
  dialect: "postgres",
  prompt: "Revenue by plan for the last two months",
  mode: "sql-and-results",
  sourceId: process.env.KAIZEN_SOURCE_ID!,
  guardrails: { readOnly: true, allowTables: ["invoices", "plans"] },
});

console.log(result.sql);
console.table(result.rows ?? []);

Set KAIZEN_API_KEY and KAIZEN_SOURCE_ID. The sql-and-results mode runs the generated query against that persisted source. Node 18 or newer, as ESM.

Start with a question.
Build what comes next.

Create an account to use Akuma, Enzan and Sōzō through Kaizen Platform today. Kai is coming soon and is not part of an account yet.