What actually breaks in RAG
The demo works on three documents. Production breaks in ways that have nothing to do with the model.
Aug 2026 · 7 MIN READ
#ai#rag#llm#retrieval
POSTS
Notes from the systems
Embeddings are not understanding
A vector is a point in space, not a mind. Know what similarity actually captures before you build a search system on it.
Aug 2026 · 6 MIN READ
#ai#embeddings#postgres#search
The word "agent" is hiding a decision
Most 'agents' are just function calls in a loop. Know the difference before you pick a framework, or you will pay for it in debugging.
Aug 2026 · 6 MIN READ
#ai#agents#llm#architecture
Stop asking the model for JSON
The model will give you JSON, and then it will lie about it. Parse, validate, repair, and retry like you mean it.
Aug 2026 · 6 MIN READ
#ai#llm#api#testing
Your search problem is not solved by vectors
Vector search finds meaning and misses the word. Keyword search finds the word and misses the meaning. Production search needs both.
Aug 2026 · 6 MIN READ
#search#ai#postgres#retrieval
LLM features need a test suite, not vibes
You would not ship a refactor without tests. You will ship a model change without evals, and then wonder why the quality vanished.
Aug 2026 · 6 MIN READ
#ai#llm#testing#evals
The boring skill that compounds: reading code
Everyone trains for writing code. Nobody trains for reading it. The developers who read well get unblocked first, and it is not close.
Aug 2026 · 6 MIN READ
#career#code-reading#collaboration#debugging
Slow tests are a debt trap
A test suite that takes forty minutes does not get run. A suite that is not run is not a suite. Speed is a correctness feature.
Aug 2026 · 6 MIN READ
#testing#ci#developer-experience#devops
Code review is a conversation
The best review comments ask questions instead of giving orders. The best reviewers treat the author as a collaborator, not an error report.
Aug 2026 · 6 MIN READ
#code-review#collaboration#culture#career
Shipping Terraform IaC without fear
Plan reviews, state discipline, and the small habits that make infrastructure changes boring.
Aug 2026 · 1 MIN READ
#terraform#iac#devops
The rewrite math
The rule everyone quotes, 'never rewrite', is a rule about systems that are still paying for themselves. Do the math, then decide.
Aug 2026 · 6 MIN READ
#architecture#refactoring#career#decision-making
TypeScript strict mode is a deal with your future self
Strict mode costs you a week of fighting the compiler now and pays you back every day after. The teams that skip it never catch up.
Jul 2026 · 6 MIN READ
#typescript#frontend#javascript#engineering-practice
Expand and contract: the migration that does not break things
Change the schema, the code, and the deploy separately, and you can migrate a live database without a maintenance window or a prayer.
Jul 2026 · 6 MIN READ
#databases#migrations#postgres#architecture
Idempotency keys: the silent contract between your API and the world
The network can duplicate anything, and your API cannot tell a retry from a second request. Idempotency keys are how you keep the books honest.
Jul 2026 · 6 MIN READ
#api#payments#distributed-systems#reliability
Webhooks are just HTTP, and that is why they keep breaking
A webhook is a POST request your server did not ask for. Everything else, signatures, retries, ordering, is the work you have to do to make it safe.
Jul 2026 · 7 MIN READ
#webhooks#payments#api#reliability
Isolation levels: where databases get honest
The database will not tell you when your reads are lying. It will just give you the settings and let you find out at 3am.
Jul 2026 · 7 MIN READ
#databases#postgres#distributed-systems#backend
Rate limiting in practice
The 429 is the politest word in the API vocabulary. The algorithms, the headers, and the decisions you make before the traffic arrives.
Jul 2026 · 7 MIN READ
#api#rate-limiting#backend#reliability
Redis is more than a cache
The same server that speeds up your reads can be a queue, a lock, a counter, a session store, and a rate limiter. Just do not store your source of truth in it.
Jul 2026 · 7 MIN READ
#redis#databases#backend#architecture
The lies we tell about staging
Staging is supposed to be production with training wheels. Usually it is production's neglected cousin, and the difference is where the incidents come from.
Jul 2026 · 6 MIN READ
#devops#environments#reliability#ci-cd
Containers vs serverless: the honest tradeoff
Containers give you control and a bill. Serverless gives you a bill and some control. Pick based on the shape of your traffic, not the hype.
Jul 2026 · 6 MIN READ
#devops#serverless#containers#architecture
Metrics, logs, and traces, in that order
When the pager goes off, metrics tell you something is wrong, logs tell you what, and traces tell you where. The teams that skip the first step find out why the other two are not enough.
Jul 2026 · 6 MIN READ
#observability#monitoring#devops#reliability
The CI/CD pipeline is a product
The pipeline is the most used tool your team has, and most teams treat it like a plumbing accident. Slow it down and you tax every developer, every commit, forever.
Jul 2026 · 7 MIN READ
#ci-cd#devops#developer-experience#automation
OAuth 2 is not a login button
That 'Sign in with Google' button is the least interesting part of the protocol. The tokens, the scopes, and the redirect URI are where the security actually lives.
Jul 2026 · 7 MIN READ
#security#oauth#auth#web
The last password hashing guide you need
MD5 was fast because nobody thought the database would leak. Every fast hash is a gift to an attacker. The slow ones are the whole point.
Jul 2026 · 8 MIN READ
#security#passwords#cryptography#backend