01
← → to navigate
Nekt Training · ~1h30

From your data to intelligence.

Six modules, from the first sync to the responding agent. Practical, with mockups, examples and real use cases.
academy.nekt.com
Agenda · six modules
The path we go traverse
01
Ingestion
Sync, streams and schedule, without waste.
02
Storage
Lakehouse, catalog and layers Bronze, Silver, Gold.
03
Context
The semantic layer that makes the AI get it right.
04
Transformations
SQL queries and Python notebooks.
05
Activation
MCP, Data API, destinations and BI.
06
Governance
Observability, alerts and permissions.
00
BEFORE YOU START

The model mental

The data maturity journey, and why value comes before optimization.
The root of the problem
Each tool is a island
Today · scattered data
HSHubSpot
MMeta Ads
GGoogle Ads
$ERP
GSSheets
PGPostgres
→
With Nekt · one place
Unique environment
Lakehouse + Catalog
gold.receita_por_canal gold.funil_por_origem gold.clientes_360
Nekt connects data from all your tools in a single environment. Without this, each source tells a different story.
What fragmentation causes
The same number, four answers What was the revenue for March?
M
Marketing
ad panel
R$ 205 thousand
Revenue attributed to campaigns
V
Sales
CRM
R$ 188 thousand
Sum of deals marked as won
D
Board
manual spreadsheet
R$ 176 thousand
Hand consolidated closure
F
Financial
ERP
R$ 152 thousand
Notes issued and reconciled
152
176
188
205
R$53 thousand difference Four systems, four truths, and no one closes the number. Without a single basis, data generates a meeting, not a decision.
Maturity Model
Four steps, value from an early age
01
Consolidate
All sources together in Nekt's Lakehouse.
02
Extract value
Context plus AI, and you're already talking to the data. The value appears here, before any modeling.
03
Transform
Model queries that have become routine, when it pays off.
04
Production
Apps and agents running in the team's daily life.
The principle that guides everything
Context and agents already deliver assertiveness that surprises. The transformation comes later, when it pays off.
Optimization is not the beginning of the journey, it is the consequence of finding what matters. The six modules follow this order.
01
MODULE 01

Ingestion of data

Bringing the right data, in the right way, at the right time, without surprise costs.
What can you connect
Structured and unstructured
HS
HubSpot
Deals, contacts and CRM pipeline
M
Meta Ads
Campaigns, expenditure and results
G
Google Ads
Clicks, cost and conversions
PG
Postgres
Tables directly from the schema
SQL
MySQL
Product transactional banking
GS
Google Sheets
Spreadsheets and manual controls
S3
S3/Files
CSV, Excel and Parquet
API
REST API
Any source with endpoint
Structured (APIs, databases, spreadsheets) and unstructured (PDFs, images, docs).
Growth · Custom

Bringing unstructured data into the Catalog, PDFs, images and documents for AI to read, is a feature of the Growth and Custom plans.

Start with the question
From question to source
Business question
Which channel brings in the most revenue?
→
Required data
Sales + lead origin
→
Source
E-commerce + CRM
Reverse engineering: start from what the business asks, not from the source.
Streams · what to bring from the source
Not every table needs come
source · HubSpot streams
✓dealssales pipeline
✓contactswho is the lead
✓companieslead company
○email_eventsno need now
○marketing_emailsno need now
Bring only what you need

Each stream is one more table to synchronize. Fewer streams, faster sync and lean catalog. Mark what answers a question and turn on the rest when you need it.

Sync and cost
Update x cost
Full Sync
Full
Rereads the entire source with each run. Simple and error-proof: the destination is always an exact portrait of the origin.
Incremental Sync
Incremental
Just what has changed since the last sync. Cheap on large table, needs reliable update column.
Full · rereads the entire table
5M lines reread every sync · minutes to hours
Incremental · only what changed
~2 thousand lines played · seconds
Same table, same freshness. Incremental does the same job by touching only what has changed.
Scheduling and triggers
Frequency at the source, the rest by event
Source · scheduled
1x per day, business days
→
Query · by event
runs when the source ends
→
Notebook · by event
runs when the query ends
→
Consumption
panel ready at 8am
Chain by event

Mark the frequency only at the source. The transformations run in a cascade, when the previous step ends, without random execution and without half data.

real case
A base came out of 160+ runs per week (full hourly, 24/7) for 5, incremental at night on weekdays, keeping the dashboard ready in the morning.
02
MODULE 02

Storage, catalog and layers

The lakehouse: where data lives and how it is refined, from raw to ready.
medallion architecture
Bronze, Silver and Gold
Bronze
Raw layer
How it came from the source
  • Faithful mirror of the origin
  • No editing, immutable
  • Raw type, often text
  • One set per source
refines
Silver
Treated layer
Clean and typed
  • Renamed and clear columns
  • Correct types, no duplicates
  • One table per entity
  • Pleasant to read
refines
Gold
Consumption layer
Ready for decision
  • Combine sources, apply business rules
  • Tables ready for analysis
  • This is where BI, MCP and API read from
  • A source of truth
A recommendation, not a rule

Each company organizes the layers as it makes sense. What matters is the concept: raw, refined, ready.

The same line, refined
From raw at the ready
order · bronze how did it come
{ "id": "a83f", "vl": "1234.50", "st": "APV", "cost": "c_5521" } // text, cryptic codes, difficult to use
gold.pedidos ready
order value status customer channel
a83f R$ 1,234.50 approved Silva Store Instagram
b91c R$ 890.00 pending Cafe Aurora Google
Typed, legible, with the name of the customer and channel already included. Ready to add, filter and ask questions.
The Catalog
Where everything is organized
workspace catalog layers
── bronze/ # how did it come from the source hubspot_deals meta_ads_insights ──silver/ # clean and typed deals advertisements ──gold/ # ready for consumption revenue_per_channel funnel_by_origin
  • Layer as a prefix: bronze, silver, gold give the reading of maturity at a glance.
  • Explainable name: receita_por_canal tells what it is without opening the table.
  • Good taxonomy is what makes human and IA find the right table.
The name matters more than the layer

The layer is convention. What guides consumption is the table having a clear name and being in the expected place.

03
MODULE 03

Context and layer semantics

What makes AI truly understand your data. The biggest leap in assertiveness.
Two houses, two roles
Where does each live? thing
Description in the Catalog · what the data is
table subscriptions
ColumnDescription
status int 1 = active, 2 = cancelled, 3 = overdue
mrr decimal Monthly recurring revenue, in BRL
cliente_id int Reference to the customer who owns the subscription
Context document · what the metric means
client-acme.md Context document
# Acme Corp ## Active customer Subscription with paid and live status minus 1 payment. Trials canceled they don't count. ## revenue Recognized from the date of financial cut.
Why context matters
The three ways AI err
01 · Ambiguity
Concept → entity
There are forty plausible fields for "revenue". Without context, the agent doesn't know which one is right.
02 · Outdated
Given grows old
Definitions and schema change. What the agent "knew" becomes old and he responds to the past.
03 · Recovery
No think
The information exists, but it disappears in the middle of everything. The agent does not find the right table.
Column description, context documents and semantic layer attack all three at once.
Failure framework described by Anthropic when automating its own data analysis with Claude.
External benchmark · Anthropic
What context did for itself Anthropic
95%
of business analysis questions answered by AI, without an analyst in the middle.
90%
of data PRs already update documentation written for AI.
Correct answers
21%
just the raw data
→
95%
+ semantic layer and skills
Source: Anthropic, "How Anthropic enables self-service data analytics with Claude." The semantic layer becomes the deterministic source of truth for the metrics.
The message
What drives assertiveness the most is not modeling everything. It is give context.
04
MODULE 04

Transformations of data

SQL queries and Python notebooks: modeling what has become routine, when it pays off.
Two transformation tools
Query or notebook?
SQL · declarative
Query
Joins, aggregations, filters: the day to day transformation. Scheduled run, it's quick to write and the MCP already generates it for you.
SELECT channel, SUM(value) FROM silver.pedidos GROUP BY channel
Python/PySpark · procedural
Notebook
When logic doesn't fit in SQL: step by step, libraries and integrations. Also generated by MCP in natural language.
import nekt df = nekt.load_table("Silver", "leads") df = enrich_via_api(df) nekt.save_table(df, "Gold", ...)
What only the notebook does

Enrich via external API, apply an ML model, process free text or nested JSON, and iterative calculations. Rule of thumb: SQL for most, notebook when you need logic or libs.

Example · transformation in SQL
A query that turns Gold
revenue_by_canal.sql query
CREATE TABLE gold.receita_por_canal AS SELECT p.canal, date_trunc('month', p.data) AS month, SUM(p.value) AS revenue FROM silver.pedidos p WHERE p.status = 'approved' GROUP BY 1, 2
Adds, filters and adds Silver. Run scheduled and become a ready table.
gold.receita_por_canal result
channel month revenue
Instagram2026-06R$42,000
Google2026-06R$31,500
Organic2026-06R$ 18,900
Generated by natural language

"Add the approved revenue by channel and month" and MCP writes this query for you.

Example · transformation into notebook
When SQL does not enough
enrich_leads.ipynb python
import nekt, requests # reads the treated table from the Catalog df = nekt.load_table("Silver", "leads") # enrich each lead via external API def size(cnpj): r = requests.get(f"https://api.receita/{cnpj}") return r.json()["port"] df["port"] = df["cnpj"].apply(port) # saves back to Gold tier nekt.save_table(df, "Gold", "enriched_leads")
Only the notebook does
Call one External API per line, apply a ML model, read free text or Nested JSON, and calculations that require a loop. Logic that SQL does not express.
Same destination

Query and notebook write to the same Gold. Those who consume it do not know (nor do they need to know) how the table was made.

Value first, model second
Model what repeat
  • Value first. Model only after the answer has proven that it works.
  • When a query becomes routine, then it's worth turning a table.
  • For AI, Nekt tends to favor wide tables, without joins.
Beware of over-engineering

Don't create a star schema when a wide table resolves. A beautiful model in the diagram is often expensive in practice.

Choose by consumption
Three ways of model
Dimensional · sales_fact with FKs
gold.fato_vendaskeys, you need to add
data_id
→ dim_data		 client_id
→ dim_cliente		 product_id
→ dim_produto		 value
d_031c_5521p_88R$ 1,234
d_031c_7702p_88R$890
Just keep the keys. To read customer or product, you need joins with the dimensions.
Wide table · everything in one table
gold.vendas_wide1 line per sale
date customer product channel value
06-01Silva StorePro PlanInstagramR$ 1,234
06-01Cafe AuroraLite PlanGoogleR$890
No joins, one line per event. Great for AI and exploration.
The third is consolidation: joins different systems using a common key. And Bronze Silver Gold It is a recommendation, not a rule.
05
MODULE 05

Activation and consumption

The last kilometer: where data becomes decision, action and product.
Activation overview
Four ways to consume
Natural language
MCP
AI agents query and explore data. Reads the semantic layer and responds in natural language.
Programmatic
Data API
Apps and systems ask for the same table and always receive the same answer.
Reverse ETL
Destinations
The ready-to-use tables return to the team's tools: HubSpot, spreadsheets, Slack.
Dashboards
BI
Looker, Metabase, Power BI read the same tables and assemble the panels.
Four doors, one source of truth. They all read the same ready-to-consume tables.
Deterministic vs Exploratory
Data API vs. MCP
Data API · programmatic · deterministic
GET/v1/tables/gold.uso_por_conta?conta_id=c_5521
answer
{
  "account_id": "c_5521",
  "uso_mes": 1842,
  "limit": 5000
}
A system asks, the same table responds the same every time.
MCP · natural language · exploratory
Person
Why did usage drop this week? And what does the plan look like?
Agent
Reads the Semantic Layer, finds the right tables and responds in natural language, without you writing SQL.
A person asks, the agent finds the way.
The way of the question
The path from the question to the answer
QUESTION language natural AGENT AI Client via MCP GOLD LAYER ready tables wide, reliable SEMANTIC LAYER context documents, rules and definitions ANSWER "Instagram, R$42 thousand"
The question comes in Portuguese, the agent reads Gold and Semantics via MCP, the answer comes back ready.
Destinations reverse ETL
The data processed back to tools
Gold
gold.lead_score
→
Destination
sync back
→
Team tool
HubSpot · Sheets · Slack
HubSpot · contact updated field
companySilva Store
lead_score87 · hot
sizemedium
Activate where the team already works

Seller won't open Nekt's Lakehouse. The score calculated in Gold appears directly on the HubSpot card, where it already exists.

BI · dashboards
Dashboards reading the Gold
Revenue by channel · June dashboard
Instagram
R$42,000
Google
R$31,500
Organic
R$ 18,900
Email
R$ 11,700
One source, multiple consumptions

The same gold.receita_por_canal that the agent reads feeds the BI dashboard. Nobody recalculates, nobody diverges.

Connect Looker, Metabase or Power BI to Nekt's Lakehouse and the panel reads Gold directly.
06
MODULE 06

Observability and governance

Run with confidence, warn before the problem arrives and control who sees what.
When something fails
Read the log and summary by AI
run #1 · failed Runs
failure Source Meta Ads · 30/06 05:58 · 14s · 401 Unauthorized
Analyst Summary · AI-Generated
What happened
The source's credentials have expired and the source responded 401 Unauthorized. The run stopped right at the beginning and no lines were processed.
Recommended actions
1) Update credentials  ·  2) Check permissions  ·  3) Reprocess the pipeline
The first measure in the event of a failure: open the run and read. Nekt already explains what happened and what to do.
Alerts
Notify the customer beforehand perceive
#data-alerts Slack
N
Nekt · Sat 08:02
Critical pipeline failed, revenue panel did not update. Source Meta Ads, 401. Summary and action in the run #1.
No alert
Friday 11pm failed. Monday 9am customer calls asking from the panel.
With alert
Saturday 8am the team is notified and resolves in 5 minutes, before anyone noticed.
Configure on critical pipelines

Alert via Slack or email on routines that someone depends on. The cheapest thing is to warn early.

Governance and permissions
Who sees the what Growth
Level 01
Viewer
See and use, read-only. Do not edit or delete anything.
Level 02
Editor
Everything from the Viewer, and edit and create: descriptions, sources, queries and tables.
Level 03
Manager
Everything from the Editor, and delete the resource and grants access to other members.
The link with AI

The agent sees exactly what the token user sees. Token Scoped limits to the chosen tables.

07
IN PRACTICE

Playbooks by vertical

The six modules assembled end-to-end for your industry.
Panorama
A playbook for each type of activity business
Multi-tenant agency
Media and CRM by customer
Sources of each account consolidated, isolated and comparable.
SaaS/Product
Churn, retention, MRR
Cohorts, recurring revenue and base health signals.
E-commerce
Orders and assignment
Products, orders and the origin that brought each sale.
Accounting
Consolidation financial
Numbers for each customer standardized on a single basis.
Construction
Cost and advance per work
Budgeted against actual, physical and financial per site.
Real estate
Funnel and commissions
Portfolio, sales funnel and broker commission.
Scenario · gym management platform
Example: SaaS
gold.cohort_retencao retention %
COHORT M0 M1 M2 M3
Jan 100 72 58 49
Feb 100 75 61 ·
Sea 100 78 · ·
Each row is a crop of new customers, each column a month after entry.
Value SaaS
One cohort dashboard shows who stays and who disappears, month by month. And one agent responds immediately: which crop retains best and what's different about it.
Caution

Churn only counts subscriptions that have already paid at least once. A canceled trial is not churn.

Scenario · agency with 12 clients
Example: Agency
gold.consolidado_midia 1 table · N customers
date customer channel investment leads
2026-06-01 Silva Store Goal R$320 18
2026-06-01 Silva Store Google R$410 22
2026-06-01 Cafe Aurora Goal R$ 280 15
2026-06-01 Cafe Aurora Google R$ 190 9
Meta and Google for each account, in the same table, with standardized columns.
A base, several customers, without mixing.
Multi-tenant isolation

Prefix or column per customer ensures data never gets mixed up.

Get to work
Your data can now
turn value today.
Connect a source, provide context, plug in the MCP and ask. The rest is evolution, at your own pace.
academy.nekt.com docs.nekt.com
Thank you