Writing

Notes on building AI systems

Technical writing on inference, evaluation, agents, and reliability. Topics are listed honestly with their status (planned, in progress, or published), and nothing is passed off as finished before it is.

• Designing an LLM gateway you can actually operate

  Planned

  Routing, caching, quotas, and failover across providers, and why the control plane matters as much as the data plane.

  Related project: LLM Inference Gateway

• Evaluation as a CI gate, not a spreadsheet

  Planned

  Golden sets, regression thresholds, and a data flywheel that keeps model quality from silently drifting.

  Related project: Eval & Data Flywheel

• Agents that ask permission

  Planned

  Permission-aware retrieval, guarded tools, and human approval for building agentic systems that are safe to audit.

  Related project: Agentic RAG Copilot

• AI SRE: golden signals for model-serving systems

  Planned

  Latency, cost, error, and quality as first-class signals, and what a degraded mode should look like.

  Related project: LLM Inference Gateway