Field Notes
Notes on event-driven Java, Spring Boot, Kafka, and secure systems for regulated banking.
Running Anthropic's prompt-audit over 2,700 lines of Claude Code steering before moving to Fable 5.1. The grep for pressure language hit once. The cruft was drift and changelog phrasing, and seven of the eight recommended prompt snippets were already in the harness.
I had Anthropic's always-on security plugin installed for a month and assumed it was reviewing every turn. Its own log said otherwise. What each layer actually does, how far its configuration goes, and the one hook that says no.
LLMs, the whole thing · Part 7
Section 3.4 of Raschka's book on six tokens with real PyTorch output: three shared learned matrices, eighteen private q/k/v vectors, why d_in is forced, why scores are divided by the square root of d_k, and why 'weights' now means two different things.
I opened a FastAPI file as a Spring engineer and kept noticing what wasn't there: no MapStruct mapper, no @Valid, no springdoc. The type signature had already done the work.
The second half of RAG: retrieve the right chunks, ground the prompt in them, and let DeepSeek answer - or honestly say 'I don't know' when the context is silent.
RAG has two halves, and only one uses a language model. A walkthrough of the other one - read, chunk, embed, index - built with Spring AI, pgvector and Ollama.
Security and privacy, in practice
A logging-only honeypot on my Hetzner box: an unprivileged user, a systemd sandbox with one non-obvious trap that silences it, and fail2ban to turn every knock into a self-expiring firewall ban.
Most 'chat with your docs' tools retrieve fragments at query time. OpenKB compiles the synthesis once, at ingest - into a Markdown wiki you can read, diff, and ship as an installable Claude Code skill.
Traces, Tempo and the LGTM stack
Wiring Spring Boot 4 + Spring Batch 6 to emit job/step traces into Tempo and spring_batch_* metrics into Prometheus — and the four config traps that each failed the same way: green build, zero errors, nothing exported.
The positional-embedding sum quietly relies on broadcasting. Here is the whole rule — align from the right, stretch the 1s — and why PyTorch never copies a byte.