Field notes
Architecture, automation, and product lessons from building business systems and governed AI on the JVM.
Agent harnesses re-send your entire transcript on every turn, and /compact destroys the KV-cache prefix that made inference fast in the first place. Here is the measured cost of both — and the deterministic, non-LLM fix we built in Rust.
Read article
TramAI is at 0.6.0 and heading to a governed AI control plane. What the 0.7 to 1.0 roadmap means for JVM teams and EU AI Act deployers.
Read article
EU AI Act deployer obligations (Article 26) apply now. What deployers must do, and how to build the evidence trail regulators ask for.
Read articleThe Digital Omnibus (2026/1744) postponed the EU AI Act high-risk obligations. What changed, what did not, and what it means for your compliance timeline.
Read articleEU AI Act compliance in 2026: SaaS vs. consulting vs. self-hosted. Compare the three models on ownership, data residency, and audit evidence.
Read article
LangChain is an orchestration framework. TramAI is a governance runtime. If you need to demonstrate EU AI Act compliance to an auditor, the difference matters. Here is a practical comparison for Java/Spring Boot teams.
Read articleLearn why production AI approval requires suspension, durable state, replay-safe continuation, idempotency, and explicit approval outcomes—not just a confirmation dialog.
Read articleWhy prompt engineering is not security: how to govern AI tool calling, side effects, durable human approval, and runtime policy in Java and Spring Boot applications.
Read articleLearn how to integrate LLMs into Java and Spring Boot using typed contracts, structured output, model routing, tool governance, human approval, resilience, and audit evidence.
Read article
A practical case study of building Sofia, an AI voice agent that calls new leads, captures structured context, and sends a meeting-booking follow-up.
Read articleThe engineering decisions behind TramAI and why production AI systems need explicit boundaries, permissions, portability, and auditability.
Read article