Building agentic AI systems and the platforms they run on.

• Agentic AI systems — production agents with retrieval, tools, governance, and eval gates.
• Agent harnesses — built on the Claude Agents SDK and the OpenAI Agents SDK, picking the right primitives for the job (subagents, tool use, structured outputs, long-running tasks).
• Data integration on Databricks — Mosaic AI Agent Framework, Vector Search, Lakeflow SDP, Unity Catalog, Lakebase. Where the agents get their grounded context.
• AI-assisted coding workflows — Claude Code and Codex as the implementation surface, Spec-Kit + constitutions as the design surface, eval suites as the merge gate.
• Container platforms — AKS (Azure) and EKS (AWS). Kubernetes-native deployment for everything that doesn't live inside Databricks.

• Agent SDKs — Claude Agents SDK, OpenAI Agents SDK, Mosaic AI Agent Framework. Comparative strengths: subagent isolation vs. handoffs, tool-use ergonomics, OBO identity, latency under realistic loads.
• Document intelligence — ai_parse_document, layout-aware section explosion, typed KPI extraction, quality rubrics that decide what reaches the index.
• AI-native developer tooling — Spec-Kit constitutions, Claude Code skill bundles, Codex orchestration, CLEARS eval gates as deploy promoters.
• Cloud-native platforms — AKS / EKS, Helm, Terraform, identity federation, secrets management. Boring, load-bearing.

PDF ─▶ Bronze ─▶ Silver
                  │
                  ▼
             Gold + quality
                  │
                  ▼
          Vector Search index
                  │
                  ▼
          Cited-answer agent
                  │
                  ▼
            Analyst chat

Agent SDKs & runtimes

AI-assisted development

Data & lakehouse

Platform & cloud

Languages

The interesting question for the next few years isn't "which agent SDK wins" or "Claude Code vs. Codex." It's: what's the smallest, most testable unit of intent we can hand an AI implementer, and what's the eval surface that decides whether their output ships? The SDKs are tactical; specifications, constitutions, and eval gates are the actual leverage.

• 💼 LinkedIn — best for collaboration / consulting
• 🐙 Open issues on any of my repos
• 🌱 Always open to: multi-SDK agent patterns, AI-assisted dev workflows, governed retrieval on Databricks, AKS/EKS at scale