Applied AI
8 min
AI in production: RAG, evals and prompt injection
Anybody can build a demo that impresses. What separates the demo from the product is three disciplines, and most teams have none of them.
ragevals
Read articleArchive · Blog
Notes on engineering, applied AI and the boring decisions nobody covers in tutorials.
Every article comes out of something I built or watched break: payments, distributed systems, AI in production. Start with whatever is biting you right now, or work down the list from the top.
24 articles
Featured
Engineering
7 min
Data and analytics: from OLTP to lakehouse
Two questions come up in every company: why did the report take down production, and why is the number on my dashboard different from yours. Both have the same root cause.
oltpolap
Read articleApplied AI
5 min
What changes when the AI agent leaves the laptop
The demo works in five minutes. What nobody shows is the evaluation, cost and failure layer that separates a prototype from something a real user can survive.
llmagents
Read articleThe whole archive
Engineering
7 min
Legacy: characterise, seam and strangle
Three techniques that let you safely change a system you did not write, do not understand, and cannot stop.
legacy-codecharacterisation-tests
Read articleEngineering
6 min
Tests worth what they cost
Tests do not exist to prove the code is right. They exist so you can change it tomorrow without fear. That change of goal reorganises everything.
testingtest-pyramid
Read articleEngineering
4 min
The idempotent webhook: the bug that only shows up in production
Every gateway redelivers webhooks. If your handler is not idempotent, you will find out on the day the same sale gets counted three times.
webhooksidempotency
Read articleEngineering
6 min
OAuth2, OIDC and JWT demystified
The three most confused acronyms in authentication, what each one solves, and the mistakes that show up in almost every codebase.
oauth2oidc
Read articleEngineering
7 min
The vulnerabilities that actually show up
Most breaches that make the news do not use sophisticated technique. They use one of these six flaws, and all six have a known, cheap fix.
securityowasp
Read articleEngineering
6 min
Diagnose the network in ten minutes
A six command routine that turns "it must be the network" into "it is layer X, on hop Y".
network-debuggingdns
Read articleEngineering
6 min
Why your service degrades over time
The service starts fine and gets worse over days. Three causes explain almost every case, and all three have distinct signatures.
memory-leakcontention
Read articleEngineering
5 min
Event loop or thread per request: how to choose
The choice between the two models is not taste and not fashion. It is determined by the profile of your load, and there is a calculation that settles it.
event-loopthread-per-request
Read articleEngineering
7 min
Concurrency: race conditions, locks and deadlocks
The bugs that do not happen on your machine, do not happen in tests, and do happen in production. How to recognise them by reading code.
concurrencyrace-conditions
Read articleEngineering
6 min
Which data structure for which problem
A decision guide organised by problem, not by structure. You have a need; which structure answers it, and at what cost.
data-structureshash
Read articleEngineering
5 min
B-tree vs LSM-tree: inside your database's index
Why Postgres is good at ranges and Cassandra is good at writes. It is not marketing: it is the tree each one picked, and the bill each choice comes with.
b-treelsm-tree
Read articleApplied AI
6 min
What an LLM really is: tokens, embeddings and attention
No mysticism and no maths: what the model does, why it hallucinates, and how that changes your architecture decisions.
llmtokens
Read articleEngineering
6 min
Observability: metrics, logs and traces, and what each one answers
The three pillars are not interchangeable. Each answers a different question, and using the wrong one is why investigations drag on for hours.
observabilitytracing
Read articleEngineering
5 min
Schema migration without downtime
Why renaming a column breaks production, and the four step pattern that turns a migration into a Tuesday afternoon task.
migrationexpand-contract
Read articleEngineering
6 min
The five resilience patterns that stop a cascading failure
How one slow secondary dependency takes down the whole system in ninety seconds, and the five patterns that prevent it.
resiliencetimeouts
Read articleEngineering
5 min
Saga and outbox: a transaction without a distributed transaction
You need to debit one account and credit another, and they live in different databases. The two patterns that solve it in practice, and what to avoid.
distributed-systemsmicroservices
Read articleEngineering
5 min
CAP, PACELC and the consistency models
CAP is the most quoted and most badly stated theorem in distributed computing. This article fixes the statement and shows the vocabulary you actually use day to day.
distributed-systemscap
Read articleEngineering
6 min
Sharding: a decision guide
Sharding is the most irreversible decision in a data system. This is the guide to deciding whether you need it and, if you do, how to pick the key.
shardingdatabases
Read articleEngineering
6 min
Indexes, EXPLAIN and the slow query
How an index works on the inside, why column order decides everything, and how to read an execution plan to know what to do.
postgresdatabases
Read articleEngineering
5 min
Caching: the four traps nobody anticipates
Adding a cache is easy. The hard part is living with the four consequences it creates, and all four have known solutions.
cacheredis
Read articleEngineering
5 min
Latency, throughput and percentiles: the guide that settles the argument
Your average latency is lying to you. The three concepts that turn 'the system is slow' into a sentence with a number, an endpoint and a percentile.
performancelatency
Read articleTalk to me
Questions about the article? Message me on WhatsApp
No form and no mailing list. If you disagree with something I wrote, or want to tell me how you solved it, the conversation goes straight to me.