Real-Time Intelligence on Confluent

01 / INNOVATION

Forecast it. Then optimize it.

A pump caught drifting today is a work order, next week it is an outage. The model that forecasts the series is also a model of the plant, so an optimizer can search the setpoints inside the guardrails.

Which setting earns the most and still meets spec?

Two knobs, three objectives: quality, energy, margin. The model forecasts what every setting does. The optimizer searches, respects the limits, and the operator approves.

View

How much we feed, t/h196

How hot we run, °C1418

Before and after

	Before	After
Feed	196 t/h	–
Temperature	1,418 °C	–
Quality	94.8%	–
Energy	89 kWh/t	–
Margin	baseline	–

1Forecast quality and energy for every setting

2Drop the ones that miss spec or cross a limit

3Pick the most profitable, send to the operator

recommendationPress Optimize

Illustrative: at 1.2 million tons a year, $2 a ton is about $2.4M a year for one mill.

2 to 5×faster optimization

93 → 99%quality accuracy

$2per ton

Design partners in cement, steel, pulp and paper, food and telecommunications

No training. Draw anything.

The model has never seen your signal. Sketch one and watch it forecast.

Draw a signal

Drag across the canvas, or start from a shape.

Shapes

Model

Horizon 48

Draw here → release → forecast

–pattern found

0training steps

1.4Mparameters

–forecast time

Same call for a sensor, a SKU, a card or a cell tower

Then something acts on it.

An agent reads the forecast and the flag, reasons, and writes the decision to a topic. Experts set the gates.

Stream to decision

One event, five hops, under a second.

Scenario

Kafka topic

pos_sales

waiting for an event

AI_FORECAST

Granite TTM

AI_DETECT_ANOMALIES

99% band

Streaming Agent

AI_RUN_AGENT

Action topic

purchase_orders

// agent log. press Run one event.

0decisions

–stream to action

5expert gates

The three statements behind it

-- 1. wrap the model as a tool
CREATE TOOL forecast_sku AS
  AI_FORECAST(units_sold, event_time,
    JSON_OBJECT('model' VALUE 'ttm', 'horizon' VALUE 7));

-- 2. bind a model, a prompt and tools
CREATE AGENT replenisher WITH (
  model  = 'granite-llm',
  prompt = 'Keep service level at P90. Raise a PO when projected stock falls under lead time demand. Explain in one line.',
  tools  = ['forecast_sku', 'stock_lookup', 'raise_po']
);

-- 3. run it per event, write decisions back
INSERT INTO purchase_orders
SELECT store_id, sku, AI_RUN_AGENT('replenisher', ROW(store_id, sku, units_sold, event_time))
FROM pos_sales;

Streaming Agents syntax as previewed by Confluent. Names and clauses can change before GA.

One stream. Every decision.

Flink runs the models on every event. Decisions land as topics, tables, alerts and agent calls.

01Streams in

POS and e-commerce

Card payments

SCADA and historians

Network telemetry

Market data

Meters and IoT

120+ connectors, Schema Registry contracts

02Confluent Cloud

Kafka topicsevery series, every event, in order. Durable and replayable for audit and re-runs

Flink SQLAI_FORECAST · AI_DETECT_ANOMALIES · TS_EMBED · TS_SEARCH · TS_OPTIMIZE

Granite Time Series modelshosted by Confluent, zero-shot, no credentials or serving to manage, your data stays in your org

Streaming AgentsCREATE TOOL · CREATE AGENT · AI_RUN_AGENT, gates for experts

one statement per decision, billed as Flink compute

03Decisions out

Purchase orders topic

Holds and alerts

Tickets with a cause

Setpoint recommendations

Tableflow to Iceberg

Apps, BI and agents via MCP

consumed as events, tables or tool calls

1statement

0decisions in flight

0models to train

Nothing leaves Confluent Cloud. Models run next to the topic.

Four models. One family. 44M+ downloads.

Pick one. Every simulation follows.

02 / VALUE

Value decays by the minute.

Slide the moment you act. Watch what is left of the decision.

secondminutehourdayweek

Card frauda stolen card, first swipe

Stockouta store, a hot SKU, a weekend

Network faulta cell site, drop rate rising

Kiln driftquality sliding, energy climbing

Illustrative curves

1 syou act after

From 30 models to 2.

A cement producer, one mill. Flip it.

30models per mill

26weeks to deliver

93%accuracy

8 to 10 regression models per cement grade, 3 grades, retrained in batch. 6 months for one mill.

Real time, bespokecustom models on streams, rebuilt per use case

Autonomous enterprisefoundation models and governed agents on Confluent

Bespoke and batchwhere most teams are today

Zero-shot, batchfoundation models, agents as nightly jobs

bespoke → foundation models and agents

batch → real time

$2EBITDA per ton

$10s to 100s Mfinancial benefit

Client identities withheld

Real outcomes on real streams.

5 to 10× productivity. Every point of accuracy worth millions. Client identities withheld.

Telco · network operations production0 tickets without a cause

IBM Network Intelligence, Granite TSPulse

Telemetry in, anomalies out, agentic root cause, remediation human-in-the-loop or autonomous.

TSPulseagentic RCAMCP tools

Financial services · markets production~0%

Multivariate asset forecasts, Granite TTM

Per asset class, against the incumbent model. Zero-shot first, tuned where the backtest paid.

TTMmultivariatebacktested

Grocery · planning reference design+0 pts

Every SKU in every store, hourly

Around 100,000 SKUs across thousands of stores. Reorder point and purchase order raised in the stream from the P10 to P90 band.

AI_FORECASTP10 to P90Tableflow

03 / USE CASES

Inside the plant.

A cement line on Confluent. Sensor to setpoint and back.

From sensor to setpoint

Pick a story, or click the kiln, the mill or the preheater.

Story

Lens

Production optimization

Free lime is trending up. The quality forecast says 97.1% in 40 minutes, below the 98.5% target.

97.1%quality in spec · target 98.5

94kWh per ton · target 92

196t/h throughput

$0.00margin vs baseline, per ton

0 hdowntime avoided

Simulated line. The topics, the statements and the gates are the real ones.

Try it on your world.

One Flink SQL function each. Real models on real data in the playground.

Every SKU, every store, every hour

28 days with quantiles. The purchase order comes from the band.

Promo next week0%

Service level

Stock on hand400

Agent decision

SKUoat milk 1Lstore112lead time2 daysnext 7 days–reorder point–stock runs out–

purchase order–

AI_FORECAST, model ttm, horizon 28, maxContextSize 730. Promo applied as an exogenous driver.

+5 to 10accuracy points

100kSKUs, thousands of stores

hourlynot weekly

From a flag to the right action

Daily sales against the band, no labels. Inject a cause.

Inject

Confidence

Latest flag

day–actual–band–likely cause–

what the team doesNo flags. Weekend peaks are expected, not flagged.

AI_DETECT_ANOMALIES, confidencePercentage 99.9, context 512 days, streaming mode.

0days scored

0flagged

same daytime to action

Every card has its own normal

Amounts scored per card at authorization. Inject an attack.

Inject

Confidence

Latest flag · card ••4417

time–amount–expected band–pattern–

decision in the streamNo flags. Normal spend for this card.

AI_DETECT_ANOMALIES per card_id, confidencePercentage 99.5, rmse in the output row. Agent holds, steps up or allows.

0authorizations scored

0held

<50 msinside the auth window

Find every signal that looks like this

Embed the stream, search by shape. Sketch a pattern or pick one.

Pattern

Query

sketch a shape here

matches–best score–indexpgvector · Iceberg

semantic tag–

TS_EMBED on each window, TS_SEARCH by cosine similarity. Time series RAG for agents.

1.1Mparameters

GPU-freeembedding

04 / START

Run it on your stream.

Three moves. The first takes a minute.

Paste one statement

Any topic with a value and a timestamp.

SELECT event_time, value,
  AI_FORECAST(value, event_time,
    JSON_OBJECT('model' VALUE 'ttm', 'horizon' VALUE 12))
  OVER (PARTITION BY series_id ORDER BY event_time
        RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS f
FROM your_topic;

Call it from Flink SQL

Models hosted by Confluent. AI_FORECAST and AI_DETECT_ANOMALIES today, TS_EMBED, TS_SEARCH and TS_OPTIMIZE next.

Watch the walkthrough

Book a working session

Bring a topic and a decision. One afternoon, your data.

Build the note

Briefing note

Four lines. Copy it into an email to your IBM or Confluent contact.

Technical reference: functions, parameters, limits

SELECT event_time, load_kw,
  AI_FORECAST(load_kw, event_time,
    JSON_OBJECT('model' VALUE 'ttm', 'horizon' VALUE 12))
  OVER (PARTITION BY meter_id ORDER BY event_time
        RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS forecast
FROM meter_readings;

SELECT event_time, response_ms,
  AI_DETECT_ANOMALIES(response_ms, event_time,
    JSON_OBJECT('model' VALUE 'flowstate',
                'minContextSize' VALUE 20, 'maxContextSize' VALUE 512,
                'confidencePercentage' VALUE 99.0))
  OVER (PARTITION BY host_id ORDER BY event_time
        RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS anomaly
FROM server_metrics;

CREATE TABLE card_alerts AS
SELECT card_id, event_time, amount,
       anomaly.forecast_value AS expected, anomaly.lower_bound, anomaly.upper_bound, anomaly.rmse
FROM (
  SELECT card_id, event_time, amount,
    AI_DETECT_ANOMALIES(amount, event_time,
      JSON_OBJECT('model' VALUE 'ttm', 'confidencePercentage' VALUE 99.5))
    OVER (PARTITION BY card_id ORDER BY event_time
          RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS anomaly
  FROM card_payments)
WHERE anomaly.is_anomaly = TRUE;

SELECT s.sku, s.event_time AS as_of,
       step.`timestamp` AS forecast_day, step.q10 AS floor, step.q50 AS expected, step.q90 AS reorder_point
FROM (
  SELECT sku, event_time,
    AI_FORECAST(units_sold, event_time,
      JSON_OBJECT('model' VALUE 'patchtstfm', 'horizon' VALUE 7, 'maxContextSize' VALUE 730))
    OVER (PARTITION BY sku ORDER BY event_time
          RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS f
  FROM daily_sales) AS s
CROSS JOIN UNNEST(s.f.forecast) AS step (`timestamp`, mean, q10, q50, q90);

SELECT event_time, vibration_mm_s,
  AI_DETECT_ANOMALIES(vibration_mm_s, event_time,
    JSON_OBJECT('model' VALUE 'patchtstfm', 'maxContextSize' VALUE 2048,
                'upperBoundConfidencePercentage' VALUE 90.0,
                'lowerBoundConfidencePercentage' VALUE 99.9))
  OVER (PARTITION BY asset_id ORDER BY event_time
        RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS anomaly
FROM motor_telemetry;

Parameter	Default	Meaning
`model`	`ttm`	`ttm`, `flowstate`, `patchtstfm` or `timesfm-2.5`. Hosted by Confluent. Both functions.
`horizon`	5	Future steps to forecast. AI_FORECAST only.
`minContextSize`	20	Points required before output starts. Fields are null until then.
`maxContextSize`	512	Most recent points used. Older points are truncated.
`confidencePercentage`	99.0	Symmetric prediction interval. AI_DETECT_ANOMALIES only.
`upperBoundConfidencePercentage` / `lowerBound…`	null	Asymmetric overrides for one side of the interval.
`rmseWindowSize`	5	Rolling window for the RMSE of predictions.

Model	Horizon	Context	Notes
`ttm`	up to 96	52 to 512	Picks the TTM variant that matches horizon and context. Fast default.
`flowstate`	over 6	up to 4,096	Context plus horizon under 4,102. Seconds-level and hourly data alike.
`patchtstfm`	up to 8,176	16 to 8,064	Context plus horizon up to 8,192. Long horizons, full distributions.
`timesfm-2.5`	model limit	multiple of 32	Google Research model, also available in both functions.

Output of AI_FORECAST: forecast ARRAY<ROW<timestamp, mean, q10, q50, q90>> and metadata. Output of AI_DETECT_ANOMALIES: is_anomaly, forecast_value, lower_bound, upper_bound, actual_value, timestamp, rmse. Inputs: value DOUBLE, timestamp TIMESTAMP(3), evenly spaced for best results. Billed as Flink compute.

AI_FORECAST reference AI_DETECT_ANOMALIES reference Confluent quickstart repo Granite models on Hugging Face granite-tsfm on GitHub GIFT-Eval leaderboard TSPulse, IBM Research Real-time intelligence, IBM on Hugging Face IBM announcement Confluent Intelligence update

Predict the next stockout
Optimize for service level

Live stream

Forecast it. Then optimize it.

Which setting earns the most and still meets spec?

No training. Draw anything.

Draw a signal

Then something acts on it.

Stream to decision

One stream. Every decision.

Four models. One family. 44M+ downloads.

TTM

FlowState

PatchTST-FM

TSPulse

Value decays by the minute.

From 30 models to 2.

Real outcomes on real streams.

Inside the plant.

From sensor to setpoint

Try it on your world.

Every SKU, every store, every hour

From a flag to the right action

Every card has its own normal

Find every signal that looks like this

Run it on your stream.

Paste one statement

Call it from Flink SQL

Book a working session

Briefing note

Real-time forecasting on Confluent Cloud, in the console