// Solutions · Analytics

Every line, every shift, one pane of glass.

The edge devices contextualize the signal and the private cloud keeps it alive. This layer is the payoff: dashboards, reports, and models that read four connected engines through one semantic layer, with one definition of every metric. A shift report lands in milliseconds, a three-year question sweeps 41.8 billion rows in seconds, and the person watching the screen never learns which engine did the work.

dt-analytics · plant-03 · packaging LIVE
OEE, this shift
0% 1.8 pts
Throughput, line 3
0u/h 3.4%
First-pass yield
0% 0.2 pts
Downtime, today
0min 18%
Throughput · line 3 · last 24 hunits/hr
OEE breakdownshift
Availability92.1%
Performance92.7%
Quality98.6%
Line status
LINE 1 · fillerRUN
LINE 2 · fillerRUN
LINE 3 · packerRUN
LINE 4 · capperDOWN 6 min
LINE 5 · mixerCIP
Downtime by cause · 7 daysminutes
Energy · site · todaykWh by hour
// illustrative data · every number on this board traces to a source tag, an asset model version & a metric definition
// Connected databases01

Four engines feed that screen.

Manufacturing data refuses to live in one database, and it shouldn't: transactions, telemetry, batch files, and business context each want a different engine. So each gets the right one, and an analytics layer on top joins them, holds the metric definitions, and serves every dashboard, report, and model from the same catalog. OEE means one thing here, in every plant, in every meeting.

MSSQL

The transactional record: orders, confirmations, quality results, e-signatures. Row-exact and constraint-checked.

4 ms lookups · audit-ready
CLICKHOUSE

The telemetry columnstore: every sensor value from every plant, compressed ~10:1 and replicated across nodes.

41.8B rows · seconds
S3 LAKE

Batch records, lab exports, legacy CSV drops, and cold history, queryable in place without loading anything first.

versioned · query-in-place
ERP MIRROR

Items, BOMs, standards, and costs on a read-only mirror, so plant math can become money math safely.

read-only · business context
SEMANTIC LAYER
metric models · one definition of OEE federated query · joins across engines governance · roles, lineage, cache
DASHBOARDSGrafana · live lines REPORTSshift · batch · audit SQL & BIPower BI · notebooks ML MODELSpredict · detect
// Row questions, column questions02

The report and the sweep run on different physics.

"Which operator signed off order 550123?" wants one exact row, transactionally guaranteed. "What was the p95 cycle time on every filler across three years?" wants to scan billions of rows and come back before the meeting moves on. One engine cannot be best at both, so the platform runs one of each and routes every question to the engine built for it.

mssql · the row question
-- this morning's production, row by row,
-- transactionally exact, mid-shift
SELECT order_no, asset, good_count,
       reject_count, signed_by
FROM   production_confirmation
WHERE  shift_date = '2026-07-22';
1,214 rows4 ms
clickhouse · the column question
-- three years, every plant,
-- one columnstore sweep of 41.8B rows
SELECT asset,
       quantile(0.95)(cycle_time_s) AS p95
FROM   telemetry
WHERE  ts > now() - INTERVAL 3 YEAR
GROUP BY asset;
38 GB scanned compressed2.1 s
01

Transactional reports

row-exact · audit-ready

Shift summaries, order confirmations, genealogy, and e-signature audit trails come out of MSSQL exactly as they were written: row-exact, constraint-checked, and queryable mid-shift without blocking production writes. When an auditor asks, the answer is a query away.

02

Big-data sweeps

billions of rows · seconds

ClickHouse holds every telemetry point from every plant in a compressed, replicated columnstore. Year-over-year OEE, seasonal drift, golden-batch comparisons across sites: aggregate queries over billions of rows return in seconds, while ingest continues at full rate.

03

Joins across engines

orders ⋈ telemetry

The interesting questions span both: reject counts from SQL joined to the temperature trace from the columnstore, for the same order, on one chart. The analytics layer does the stitching, pushing each part of the query down to the engine that is fastest at it.

04

Live and history, one chart

stream + history · seamless

Dashboards read the Kafka tail for the last few seconds and the columnstore for everything before that, seamlessly. A line chart scrolls from three years ago to four seconds ago, and the operator watching it never learns where one engine ends and the other begins.

// The analytics themselves03

What manufacturing actually asks for.

The stack exists to answer a short list of questions every plant manager already has. These are the ones we build first, from the same connected engines, with numbers that agree with each other because they share a catalog.

OEE · plant-03 · 12 weeks%

OEE & downtime

Availability, performance, and quality from real event data with downtime reasons attached at the edge. The Pareto chart names the worst actor by line and shift, and the trend shows whether last month's fix actually moved the number.

A × P × Q · by line · by shift
SPC · fill weight · X̄ chartgrams

SPC & quality

Control charts on critical parameters, straight from the contextualized stream: X̄-R, Cpk, run rules, with violations alarmed while the batch is still in the tank. Quality history joins genealogy, so an excursion traces to its lot in minutes.

Cpk · run rules · live
Scheduled output · today06:00 run
shift_report_2026-07-24_night.pdf06:02
batch_summary_B-88121.pdf06:02
genealogy_packet_LOT-2214.pdf06:03
downtime_pareto_week30.xlsx06:03
compliance_packet_july.pdfSCHEDULED18:00

Reports that write themselves

Shift handoffs, batch summaries, and compliance packets are generated on schedule from the transactional record, signed data included. The report a supervisor used to assemble from three systems and a clipboard arrives by email before the shift meeting.

scheduled · signed · delivered
Anomaly · pump P-204 vibrationmm/s RMS

Models that look ahead

With years of labeled history in the columnstore, prediction gets practical: anomaly detection on vibration and cycle-time signatures, remaining-useful-life on pumps and drives, soft sensors between lab samples. Trained on the cluster, scored at the edge when latency matters.

train central · score at edge
// what the customer gets

The dashboards are the visible part. What you're actually buying is agreement:

  • One number per question. OEE, yield, and downtime are defined once in the metric catalog, so the Monday meeting argues about causes rather than calculations.
  • Answers at the speed of the question. Row lookups in milliseconds, three-year sweeps in seconds, live lines with sub-minute freshness, each from the engine built for it.
  • Cross-plant comparability for free. Because the edge modeled every site the same way, a dashboard built for plant 1 works on plant 12 the day its data starts flowing.
  • Self-serve for the people who ask. Engineers query with SQL, analysts with Power BI, operators with Grafana, all against the same governed catalog, behind SSO and role-based access.
  • A path from hindsight to foresight. The same history that powers the reports trains the models, so predictive maintenance is a next step on infrastructure you already have.
// Contact04

Which panel would you watch first?

Tell us the number you can't get today: the OEE you don't trust, the report that takes a day to build, the failure you'd love to see coming. We'll show you the path from your data to that screen.

Start a conversation