I put agents into production and make them pay for their run cost.
Twelve years shipping production JVM systems, the last three architecting AI systems inside a regulated healthcare-IT enterprise: multi-agent LangGraph orchestration with tool contracts and bounded repair loops, a client-side LLM gateway with multi-provider routing and deterministic token budgets, RAG on Qdrant, and the CI/CD platform underneath it.
Accuracy tells you an agent can do the job. Token economics tell you whether you should let it.
the-prompt-pipeline —
13-agent production pipeline that wrote, scored, translated and typeset a
29-chapter novel across three language editions as validated EPUB 3.3.
Six weighted scoring dimensions against a 0.95 bar, deliberate model tiering
(judgment stages on the strongest model, mechanical assembly on a cheaper one),
and an evidence/ bundle that includes an audit reporting where the output
fell short. The rubrics are real and the audit is not flattering.
audiobook-gen — EPUB, PDF and plain text into ACX-compliant audiobooks on open-weight neural TTS. A Director layer decides dialogue, emotion, prosody and pauses before anything is synthesized; a pluggable Performer layer runs the model; a Producer layer masters to spec. 558 tests, CUDA and ROCm images, and measured throughput: 8.3× realtime on an M1 GPU, 3.5× on CPU alone, so a finished audio hour costs 7.3 minutes of compute and nothing in per-character billing.
Java 17-25 · Spring Boot · Python · TypeScript · LangGraph · MCP · Qdrant · Kafka · PostgreSQL · Kubernetes · Helm · ArgoCD · GitLab CI · Prometheus/Grafana · GraalVM native-image