RAG that doesn't hallucinate
Grounding an assistant in your own documents — with citations, and an honest “not found” instead of a confident guess.
Practical writing from the engineers who build our client work — AI in production, SaaS architecture, integrations, and security. The decisions and trade-offs, not press releases. Written to be useful to the people who actually have to build the thing.
Most engineering writing is marketing. Ours is the stuff we wish we'd read.
The internet is full of tutorials that stop at the happy path and launch posts that skip the part where it broke. We write the other thing: what we learned shipping real systems for real clients — the decision behind a piece of architecture, the trade-off nobody mentions until it bites, the boring detail that turned out to matter most. Every post is grounded in something we actually built or debugged, written for the person who has to make the same call next week. No press releases, no thought-leadership filler — just the notes we'd have wanted before we started.
The demo always works. Then real users arrive with messy inputs, the latency budget tightens, costs creep, and the model confidently invents an answer no one can trace. This is the field guide we give ourselves before turning a clever prototype into a feature people can actually rely on — grounding, evaluation, guardrails, fallbacks, and the unglamorous plumbing that decides whether it survives contact with production.
Read the postAssistants, RAG over private docs, evaluation and guardrails — getting models from demo to dependable.
Multi-tenancy, billing, and the foundations a product can grow on without accumulating security debt.
Reliable syncs, webhooks, retries and idempotency — integrations that hold up when the other system misbehaves.
Auth, roles, audit trails, and least privilege — the things that are cheap up front and painful to retrofit.
Pipelines, metrics, and the difference between a chart that looks busy and one a team actually trusts.
Latency, query plans, and cloud bills — the unglamorous work that keeps a product fast and affordable.
The decisions, not the press release.
Every post is grounded in something we actually shipped — including the parts that broke.
Grounding an assistant in your own documents — with citations, and an honest “not found” instead of a confident guess.
Retries, idempotency, and replay — why the happy-path webhook demo is the easy 20% and the rest is where projects die.
Tenancy, roles, and isolation done right at the start — because retrofitting it after the first customer is the expensive way.
Every post comes from something we shipped or debugged — not a topic picked to chase search traffic.
We name what we'd do differently and where the obvious answer is wrong. The caveats are the useful part.
Focused on the decision you actually face, with the background linked rather than padded into the post.
The best posts start as a real question someone asked us. Tell us what you're wrestling with — even if it doesn't become a post, we'll point you at the most relevant thing we've written or built.