A Cold-Start Canary Is Not a Serving SLA
One successful cold start proves a path. It does not prove concurrent safety, cache independence, recovery, cleanup, or a production serving SLA.
All topics · 8 posts
Distributed Runtime Systems
Distributed execution, model serving, performance, capacity, and runtime debugging.
A Local LLM Router Is Not a Panel Lane Yet
A private local LLM route can pass health and quality checks and still remain only a candidate until controls, performance, environment integrity, role policy, and activation authority align.
Before Raising Reindex Concurrency, Prove the Memory Lane
Raising embedding reindex concurrency is not just a performance tweak. It becomes eligible only after endpoint identity, index shape, concurrency path, capacity limits, and final behavioral smoke all line up.
Synthetic Fanout Is Not Production Approval: A Safer Pattern for Agent-Run Distributed Probes
A bounded synthetic fanout can prove that an agent can coordinate a worker plane. It does not prove data approval, production readiness, or service integration. Keep those gates separate.
Container-First Distributed Model Serving: Treat Remote Workers as Disposable Proofs
Remote model-serving workers should start as disposable, contract-tested containers—not permanent host mutations. The useful lesson is how to separate adapter bugs, runtime substrate failures, and production-readiness gates.
Reachable Is Not Ready: Capability Matrices for Agent Runtime Offload
A read-only offload-runtime probe showed several remote lanes were reachable, but not equivalently ready. The safe next step was a capability matrix and dry-run executor contract, not a queue.
The 10-Second Session List: Why Prefiltering Before Row Build Matters in Agent Gateways
A self-hosted agent-gateway performance lesson: if a tiny session-list request builds hundreds of rich rows before filtering, limit is too late to save you.
Processing Background Shapes: From Radar Signals to Distributed ML Runtime Systems
How my PhD in radar signal processing translated into distributed ML runtime systems and backend execution work.