AI Engineer and Architect — agents, evals, and production LLM systems.
25 years of platform and SRE discipline applied to systems that think.
Credited by the Django project for reporting
CVE-2026-48588
(potential exposure of private data via cached Set-Cookie responses). Severity: low.
Official listing:
NVD / NIST.
Progressive roles across DevOps engineering, systems administration, network engineering, and infrastructure — building the deep Linux, networking, virtualization, and storage foundations that underpin modern cloud and AI platform architecture. Includes DevOps Engineer at Vets First Choice/Covetrus, Systems Engineer at EVO Payments, Senior System Administrator at MEMIC, and earlier roles in consulting, technical support, and national-scale deployments.
Self-directed research conducted alongside professional work, 2024–present. Private GPU hardware, production-pattern systems, real evaluations — not proofs of concept.
Dual-node GB10 cluster serving 30B–397B-class models, including 1M-context vision/reasoning. Currently tested and running: deepseek-v4-flash, deepseek-v4-flash-vision, glm-5.3-flash, and Qwen3.5-397B. Profile-switched serving with quantitative admission (KV / context), Prometheus/Grafana, and a hardware-specific vLLM fork.
Production multi-agent system: infrastructure ops, evaluation, project tracking, and workflow automation. Agents have real tools against live systems — not a demo chatbot.
Industry benchmarks (lm-evaluation-harness) plus custom agentic tests: tool-calling, schema/format, multi-turn, text quality. Independent LLM-as-judge. Dual-axis scorecard: quality and decode, not vanity tok/s. Grafana dashboards.
Inference routed through llm-switch across vLLM and SGLang. Canaries on the same cluster as production serving. Speculative decode / MTP and quant recipes measured against a ship bar: quality up, or quality and speed. Documented rollback to the daily ship.
Long-lived semantic memory across agent sessions, with retain/recall and downtime backfill — not a one-shot RAG demo.
Preference alignment and parameter-efficient fine-tuning on hybrid architectures. Pre/post eval for catastrophic forgetting. Internal guidelines for small-active-parameter MoE.
Production-pattern RAG stack — FastAPI + pgvector + Redis + Celery
PythonvLLM fork with DGX Spark inference patches
PythonAgentic automation for infrastructure management
PythonCUJ framework reference for production SRE
Markdown