Field Notes
Notes on event-driven Java, Spring Boot, Kafka, and secure systems for regulated banking.
The multi-part series and running themes, rather than one long pile.
Working through large language models from the ground up — turning text into tokens, embeddings and the geometry of meaning, the transformer architecture, attention, how these models are trained, and what separates one generation from the next.
Learning Embabel by building with it, in order — planning, durable sessions, testing without a model, tools, human-in-the-loop, tracing, exposing it over MCP.
Build it, attack it, keep extending it — the Ask Zakaria concierge from the first prompt through prompt-injection red-teaming to reading a job description and taking questions by voice.
Instrument it, learn to query it, then point it at something that thinks.
An agent as a collaborator rather than an autocomplete — design systems, durable knowledge, automation ladders, subagent fleets on my own source, and now in CI.
Attacking my own features and reading the protocols underneath them — secrets delivery, response headers, a published address for bad news, a cookie banner that sets no cookies, and what an auth flow is actually protecting you from.
Shipping and operating my own product in public — one jar serving two hosts, making it findable, dropping experiments into it, shipping a free CV builder off the side.
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.
I ran the /eli5 skill over a Jinja2 learning session and got back a self-contained, illustrated artifact. Here is what the skill is, why writing the five-year-old version sharpens your own understanding, and where it fits a learn-then-teach workflow.
The same engine powers Flask pages, Ansible playbooks and static sites. Here is Jinja2 learned from the engine out - delimiters, autoescaping, inheritance and the Environment - with every concept run by hand, coming from the JVM.
I ran a full LoRA fine-tune of Gemma 3 270M on my AMD Strix Halo iGPU. It froze the whole machine twice before three lines of config fixed it — here's the run, the fix, and what it taught a backend dev about how LLMs actually work.
I'm a Spring/Kafka backend engineer who could read Python but never wrote it. The AI world pulled me in, so I learned it by building a 10-lesson course that maps every concept back to Java, then built the tutor agent that taught me.
LLMs, the whole thing · Part 6
The first attention mechanism in Raschka's LLM book, explained the way it finally clicked for me, a Java dev who bounces off dense maths: self-attention is just three operations, a dot product, a softmax, and a weighted average.
The maths under positional embeddings, from zero: a token becomes what it is plus where it sits, and that 'plus' is a literal element-wise sum of two vectors. Why the differently-sized tables never actually clash, worked by hand, with slides.
LLMs, the whole thing · Part 5
The finale of chapter 2: how meaningless token IDs become the tensor attention actually consumes. It comes down to two lookup tables and one addition - a learnable embedding table for meaning, and a second, position-indexed table for order.
LLMs, the whole thing · Part 4
I printed my first training batch, saw shape [8, 4], and could not say what the two numbers counted. Untangling it - one short story sliced into 1,286 windows, grabbed 8 at a time - made the whole GPT input pipeline finally go transparent.