Robust pipelines start before the tool.

• 25 Jul
  I crossed 5 domains of my knowledge. AI found 130 bridges I never saw. Tuesday the why, Thursday the engine. Today what it found crossing finance, AI, method, master's and career at once: 130 connections I never saw.
  segundo-cerebro
• 23 Jul
  Why your RAG lies with confidence (and the 3 pieces that fix it) On Tuesday I showed why mining beats hoarding. Here is the jackhammer from the inside: 111 typed links, rank fusion and a grading node that refuses to lie.
  segundo-cerebro
• 21 Jul
  6 years of Zettelkasten: the day AI showed me what I already knew I wrote the notes. The AI found the connections between memories I had forgotten.
  segundo-cerebro
• 18 Jul
  IA Foundations 07: a 1M token window solves nothing if the information lands in the middle Lost in the middle is not a bug, it is an attention pattern. Position beats context size, and a critical rule on line 300 is accuracy thrown away.
  ia
• 16 Jul
  Staff Engineer is not 'Senior Senior'. It is a different mental model of impact. Senior solves what was assigned. Staff decides what nobody has been assigned yet. It is a promotion in scope, not in seniority.
  engenharia-de-dados
• 14 Jul
  Iceberg in production: 3 anti-patterns that kill your data lake in 3 months Naive partitioning, bloated manifests, forgotten compaction. How that turns into 100k GET requests on a query that should cost cents.
  engenharia-de-dados
• 11 Jul
  Your LLM agrees with what you want to hear. And that is an architecture bug, not a prompt bug Sycophancy is automated confirmation bias. You fix it with architecture: independent models, confidence-weighted consensus and a primary source that vetoes the rest.
  ia
• 09 Jul
  Green build, production on fire: when the test stub teaches a lie about the real lib A frontier model scored low on a coding benchmark with a 100% green CI. The bug was shielded by its own mock. A green test does not prove correct integration.
  ferramentas
• 04 Jul
  Zero to Expert Ep 04: Wrong partitioning costs more than not partitioning Almost everyone learns "partition to make it fast". The wrong granularity makes your pipeline slower, pricier and more fragile than no partition at all.
  zero-to-expert
• 04 Jul
  Your agent is degrading in silence. I built a 9-eval suite to prove it Measuring only the final output hides regression. Without per-step, per-role evals, agents get worse unseen. The 9 golden cases that stop it.
  ia
• 02 Jul
  Stop writing prompts by hand: DSPy and TextGrad treat prompts as compilable code Improving a prompt is not manual copywriting, it is compilation against a metric. Two Stanford frameworks prove it, and one landed in Nature.
  ferramentas
• 30 Jun
  You're running everything on the most expensive model: routing by role cuts 60-70% of your agent bill In an agent swarm, only irreversible decisions justify the frontier model. Everything else is coordination, implementation, artifact transformation.
  engenharia-de-dados
• 27 Jun
  Zero to Expert Ep 03: Idempotency, why running the pipeline twice cannot change the result Ep 01 covered data flow, Ep 02 covered DAG dependencies. Now the property that separates a reliable pipeline from a time bomb: idempotency.
  zero-to-expert
• 27 Jun
  The 5 subsystems that separate an agent that works from one that only looks like it works Last week I wrote that the LLM does nothing, the harness does. This post opens the hood: the five subsystems every working agent needs.
  ia
• 25 Jun
  You don't need a vector DB: the honest verdict on RAG in the era of 1M-token context Two retrieval systems in production with zero embeddings: when grep plus long context beats classic RAG, and when a vector DB still wins.
  ferramentas
• 23 Jun
  The producer renamed a column and nobody told us: why a pipeline without a data contract breaks in the dark There is one kind of incident I learned to fear more than a failing pipeline: the one that does NOT fail. Green job, wrong numbers for three weeks.
  engenharia-de-dados
• 20 Jun
  Why asking AI to review its own code does not work (and what does) IA Foundations 03: asking the same model to self-correct can degrade reasoning. The multi-agent alternative reaches 82.6% on HumanEval.
  ia
• 20 Jun
  Retroactive lineage is technical debt. Document at decision time or lose it. The Audit Envelope Pattern I learned the hard way: instrumenting later always costs more than instrumenting upfront.
  engenharia-de-dados
• 18 Jun
  AI won't kill the engineer. It will kill the coder. The distinction that will separate those who thrive from those who fall behind over the next 3 years.
  engenharia-de-dados
• 16 Jun
  Prompt injection is the SQL injection of AI. And almost nobody takes it seriously. OWASP LLM Top 10 #1 since 2023. The most dangerous attack vector in AI pipelines is still treated as a detail.
  engenharia-de-dados
• 13 Jun
  Zero to Expert · Ep 02: DAG dependencies, the law every orchestrator obeys Before picking Airflow, Dagster, Prefect or anything else, understand the graph.
  dag
• 13 Jun
  AI Foundations · 02: The LLM does nothing. The harness does. The most common mistake of 2026 is not separating the model from the harness. It costs money and security.
  llm
• 11 Jun
  Prompt caching: the one-line change that cuts 90% of LLM cost in production cache_control ephemeral, a 5-minute TTL, and how one line dropped 75% of the cost of my nightly pipeline.
  llm
• 10 Jun
  SQL is still the most important language in data engineering in 2026 Why the dev who outsources SQL to an ORM or pandas becomes the spreadsheet filler of the AI era.
  sql
• 04 Jun
  YouTube rate-limits its caption endpoint. Audio stays free. Why --sleep-subtitles 60 no longer fixes YouTube's HTTP 429 on captions, which endpoint still works, and what self-hosting with faster-whisper delivers vs paid SaaS.
  youtube
• 02 Jun
  Code Review of My Own Old Repo. Five Things I'd Change Today. I opened a two-year-old repo of mine. Found five anti-patterns in my own code. Honest review before continuing to cite it on my portfolio.
  code-review
• 30 May
  SLA, not trend: when batch, when streaming, when both The choice between batch and streaming is not cultural. It is technical: consumer SLA, cost, complexity. And almost nobody measures it before deciding.
  batch
• 30 May
  Data Flows Ep01: the concept that comes before any tool Knight Capital lost $440 million in 45 minutes because one server ran old code. Before any tool, there is a concept that separates a robust pipeline from a fragile one: data flow.
  zero-to-expert
• 24 May
  Airflow for 2 years: what I would do differently Two years of Airflow in production: finance, factory processes, marketing, analytics. List of what I would change if starting over, not doc theory.
  airflow
• 23 May
  Twenty AI concepts you need to understand in 2026 The complete visual guide to how AI works, from the basics to the frontier. Six sections, twenty concepts, no spare jargon.
  ai
• 17 May
  When the model should say 'I don't know' In 1998 LTCM lost $4.6 billion because its models trusted territory they had never seen. Twenty-seven years later my HMM was doing the same, at smaller scale. Here's what I learned to put between it and the trade.
  crypto
• 13 May
  Instrumenting lineage from scratch with Unity Catalog Unity Catalog captures lineage automatically, but the limits matter: UPDATE/DELETE/MERGE, RDDs, JDBC, multi-hop. What it covers and where most people get it wrong.
  databricks
• 12 May
  When Medallion Architecture gets in the way more than it helps Medallion became the default answer for any pipeline. Where it works, where it gets in the way, and the anti-pattern nobody talks about.
  medallion
• 02 May
  LGPD and ML models: what to do with data that has already become model weights A data subject requests deletion. You delete the row from the database. The model in production still carries the contribution from that CPF. What LGPD requires and what's actually possible.
  lgpd
• 30 Apr
  Delta Lake or Parquet? You're asking the wrong question Delta Lake doesn't compete with Parquet. It's built on top of it. The real decision is different, and getting it wrong silently corrupts data.
  delta-lake
• 23 Apr
  Sharpe Ratio -1.14 is an Engineering Win, Not a Failure Data leakage destroyed a +0.66 Sharpe ratio into -1.14. What I learned: 7-layer architecture, 202 tests, 90x better capital preservation than buy-and-hold.
  crypto
• 18 Apr
  Real data engineering content in Portuguese is rare. I'm going to help change that. Why a senior data engineer with experience across retail banking, payments fintech and a big-tech project in Silicon Valley decided to start writing, and what you'll find here.
  manifesto
• 16 Apr
  LGPD at the ingestion layer: 4 principles that change your architecture LGPD isn't a checklist at the end of the pipeline. It's a design constraint from the first byte that comes in. What that means in practice.
  lgpd

No posts yet. All