[ dispersion engine — v4.2 ]
Raw in.Resolvedout.
Spectra sits between your chaotic event stream and the people who operate on it. It ingests the noise — logs, topics, lake files, drifting schemas — and disperses it into three sharp bands your team can act on before lunch.
- throughput
- 0.0 M evt/s
- resolve p95
- 0.0 s
- trust
- SOC 2 · EU
[ sys.01 — the problem ]
Your lake
is not
a lake.
← in / chaos
- 19 Kafka topics, 4 naming conventions, 0 owners
- Parquet partitions landing late, twice, renamed
- plan_tier drifted varchar → json on a Friday
- Revenue dashboard disagrees with itself by 18%
- Three teams maintain three definitions of “active”
out / resolved →
- One intake. Every source typed, timed, owned.
- Drift quarantined before it reaches a dashboard.
- Revenue, risk, freshness — verified, with confidence.
- Each insight ships with an owner and a next step.
- “Active” defined once. Enforced everywhere.
Fig. 02 — The prism does not store your chaos. It refuses to pass it through.
[ sys.02 — proof ]
Measured
in prod,
not slides.
0M evt/s
sustained intake
peak 7.2M, no backpressure
0s p95
raw → resolved
from 40 min batch window
0%
insight confidence
verified, drift excluded
0%
join cost cut
median, first 90 days
[ sys.03 — sequence ]
Three
passes.
No noise
survives.
Numbered because order is the product: nothing can be resolved before it is ingested, nothing ships before it is verified.
01
Ingest
Connect Kafka, S3, Postgres, CDC and webhooks in an afternoon. Every byte gets a timestamp, a source and a schema fingerprint on arrival.
kafka → intake · s3 → intake · cdc → intake
02
Disperse
The prism splits the stream: normalize types, quarantine drift, join facts, kill duplicates. Chaos enters; three clean bands leave.
normalize · quarantine · join · dedupe
03
Operate
Revenue, risk and freshness land where work happens — warehouse marts, Slack, your dashboard — each with an owner and a confidence score.
mart · alert · owner · confidence
$ spectra connect kafka://prod --topics raw_* ✓ 19 topics
$ spectra watch --band revenue_7d
intake 4.81M evt/s schema fp a4:91:c2
quarantine plan_tier drift → held dashboard untouched
resolved €4.82M · conf 99.1% owner: m.okafor
$ spectra ship --to warehouse.marts ✓ fresh 4 min agowhy teams switch
“Our Monday numbers meeting went from ninety minutes of arguing about definitions to ten minutes of deciding. The prism ate the argument.”
— Data platform lead, Series D fintech
1.2B events/day · 6 sources · live since March
[ sys.04 — telemetry ]
The bands,
plotted
weekly.
Same three bands as the prism fan: revenue resolved, risk owned, freshness verified. Drawn from one pilot cohort, not a benchmark lab.
Resolve time fell from 40 minutes to 1.9 seconds
p95 raw → resolved · seconds, log scale · 24 pilot weeks
F3 Hairline Area · Pilot cohort · 2026
Throughput holds 4.8M events a second
sustained intake · M evt/s · 30 days prod
F2 Hairline Line · Prod EU-West · 2026
Every heavy join costs less after dispersion
compute cost · $/day · before / after · median customer, 90 days
F6 Paired Rungs · Billing export · 2026
[ sys.05 — compare ]
Three ways
to lie to
yourself.
Two of them are industry standard. Dimensions a platform team actually argues about.
dimension
diy lake + dbt
warehouse-native
spectra ◢
[ sys.06 — stack ]
Sits beside
your stack.
Kafka
intake · CDC
S3
lake files
Postgres
CDC outbox
Snowflake
marts in
BigQuery
marts in
dbt
tests stay green
Slack
alerts + owner
PagerDuty
freshness SLO
Fig. 03 — No rip-and-replace. Spectra reads what you already emit and writes where you already look.
[ sys.07 — faq ]
Asked by
every CTO.
Q1Do we rip out our warehouse?
No. Spectra sits beside it. Intake reads Kafka, S3 and Postgres CDC; resolved bands land back as versioned marts. Your warehouse stays the system of record — it just stops receiving garbage.
Q2Where does our data live?
Where you put it. EU or US region, single-tenant intake, SOC 2 Type II. Raw payloads are typed and fingerprinted at the edge; nothing trains shared models.
Q3Raw intake is free — where is the catch?
There is none to find. We charge per resolved event that ships with an owner and a confidence score. Noise you send us costs you nothing, which is also why we are motivated to turn it into signal.
Q4What if the pilot numbers do not hold?
You keep the audit: schema fingerprints, quarantine log, confidence report. If resolve p95 or confidence miss the agreed thresholds, the pilot costs nothing.
[ sys.08 — access ]
Feed the
noise.
Pilot in two weeks: we connect one source, resolve one band, and prove confidence on your own data. If the numbers do not hold, you keep the audit.
SOC 2 Type II · EU residency available · No rip-and-replace, sits beside your stack
- Week 1One source connected. Intake verified against your lake.
- Week 2First band resolved. Confidence report on your definitions.
- Day 30Three bands live. Owners assigned. Old dashboards retired.
- PriceUsage-based on resolved events. Raw intake is free — we only charge for signal.