I am one insight short of 100 on my-tech-profile.dev. I did not write a hundred essays. I wrote a map of what kept breaking when I tried to ship Java, Spring, and agents for real.
I am one insight short of 100 on my-tech-profile.dev/insights . I did not set out to write a hundred essays. I set out to leave a trail I could reuse: what broke in production, what I would not do twice, and which Java and agent patterns still held after the demo glow faded. As of 28 August 2026 the live catalog has 99 published pieces, all from this year, about 1,030 minutes of reading if you sat through every one. That is a terrible way to read them. This piece is the map I wish I had on the archive page. The count is a calendar. The threads are the product. What the shelf actually looks like I publish under categories because a flat list of 99 titles is a wall. The current split is uneven on purpose. The work pulled me there. Category Count AI Engineering 33 System Architecture 27 Coding Experience 13 Software Design Patterns 11 Enterprise Integration Patterns 6 Observability 3 AI Architecture 2 Other (CI/CD, branding, tokens, practices) 4 Tags tell the same story in a cruder way: Artificial intelligence, Architecture, LLMs, Java, Spring Boot. That is not a content strategy document. That is where the tickets were. May was the sprint that filled the shelf. June through August slowed into longer, more specific pieces: Spring AI 2.0 versus LangChain4j, MCP security, virtual threads on JMS, a typed PR reviewer. I would rather keep that direction than chase a posting streak. Four threads I kept writing until they stuck I am not going to recap ninety-nine titles. If you only remember four claims, remember these. Java and Spring are still the substrate. Virtual threads with JMS listeners, Spring Boot 4 versus 3, Modulith, DB-Scheduler, repository debt, HttpStatus printing. Agents did not replace the platform. They sat on it. When the model is noisy, the loop that is not a model is what you can test. Agents need gates, not vibes. A goal the agent cannot fake. Spec-driven development. AGENTS.md . Superpowers as a skill framework. A PR reviewer that returns ReviewVerdict , not an essay. Don't write prompts, run loops. Skills versus prompts. If APPROVE lives in the last sentence of a chat, you do not have a reviewer. You have another writer. RAG is a product problem, not a demo. Measure with RAGAS. Stop trusting pure retrieval for Job Fit; I went to ontology fusion on ITJobOpportunities because embeddings alone could not carry hiring evidence. Agentic RAG and memory showed up when chat was not enough for S&J Leathers and similar product-chat work. Filter junk before it hits the vector store. The index is part of the contract. Production architecture is still the job. Microservices decisions that actually fail. Hidden bottlenecks. Redis under 100ms. Camel observability. Chaos Monkey. Audit trails. Caddy. Cluster-friendly scheduling. Alerting. The AI posts only make sense if this layer is honest. I will not trade a green agent demo for a system nobody can operate. A starter pack if you are new to the archive If you land on Insights cold, do not start at May and scroll. Read five, then pick a thread. A Goal the Agent Cannot Fake . Typed exit conditions, not self-assessment. Spring AI 2.0 vs LangChain4j 1.18 . Same pipeline, disagreement only in the orchestrator. Ontology Knowledge Fusion for Job Fit . Why I stopped treating RAG as enough for hiring evidence. Seven microservices decisions . The non-AI spine. A production-ready AI PR reviewer . The latest cut of "Java owns APPROVE." After those, Superpowers, MCP security, and the coding-agent cost comparison are the natural next clicks if you live in Cursor or Claude Code. What I would not repeat I over-tagged early posts. "Artificial intelligence" on everything makes the tag useless. Some titles are longer than the insight. One even repeats AGENTS.md twice. I would cut those before I would write a new one. A few architecture posts are surveys (kinds of RAG, LLM architecture explained). Surveys age faster than a single production constraint. The pieces I still reread are the ones with a gate, a version, or a failure I actually hit. I also learned the archive is for operators, not for a personal brand category. The one personal-branding post is the outlier. The workbench tone is the brand. What 100 is for Number 100 should make the other 99 easier to enter, not harder to ignore. That is this article. What comes after is not "more of the same, faster." I want fewer, sharper notes: one stack version, one failure, one decision a teammate can take into a PR. The PR reviewer and Superpowers posts are closer to that than a twelve-minute tour of a whole domain. If you have been reading along, I am curious which thread you want next: Java platform, agent gates, RAG that you can measure, or the boring production architecture that makes the rest possible.