What we learned building it.
- AI systems
Retrieval recall is the RAG metric that actually decides quality
Most RAG systems fail at retrieval, not generation. How to measure recall@k on your own corpus, what a passing number looks like, and why prompt engineering cannot fix a retrieval problem.
8 min read - AI systems
When not to build an AI agent
Multi-agent architectures are oversold. A practical test for whether a workload needs an agent, a single model with tools, or ordinary deterministic code.
7 min read - Automation
Idempotency is the whole job
Automation that cannot safely run twice is a liability. Idempotency keys, durable execution, and decline-code-aware retries - the patterns that keep automated money movement correct.
7 min read - AI systems
pgvector or a dedicated vector database
When PostgreSQL with pgvector is the right default for RAG, what actually breaks at scale, and the specific signals that justify moving to a dedicated vector store.
6 min read - Performance
INP is the hard Core Web Vital, and it is a JavaScript problem
Interaction to Next Paint is the metric most sites fail, and image compression will not fix it. Where the long tasks come from, how to break them up, and what an INP budget looks like in practice.
7 min read
Who writes these
All of them are written by Manish Meena, co-founder and principal engineer at Pruning Labs, from systems the studio has run in production rather than from a survey of other people’s blog posts.
Why we publish these
Two reasons. The first is that most of what is written about AI systems is written by people who have not run one for eighteen months, and the difference shows up in what they think is hard.
The second is more self-interested: these notes are how prospective clients decide whether we know what we are talking about before the first call. If a note is wrong, tell us - corrections get published with the date they were made.
The work behind them is on the case studies, and the disciplines are described under services.