RESEARCH
NOTES.
Short-form technical writing, design fragments, mental models and engineering observations. Notes are intentionally dense, linked and revised over time.
Context Engineering Is Becoming Infrastructure.
Why the information environment around a model is becoming a first-class runtime concern for agentic systems.
RAG Is Not a Database.
Retrieval is a policy for assembling evidence under constraints. Treating it as a storage primitive hides the hard part.
Notes on Agent Memory Architectures.
A practical decomposition of working memory, episodic memory, semantic memory and durable state.
What Small Models Are Actually Good At.
Latency, specialization, routing, privacy boundaries and bounded tasks are often more important than raw benchmark rank.
Building Reliable Tool-Using Agents.
Model intelligence does not remove the need for state machines, validation, retry policies and observability.
Inference-Time Compute as a Product Primitive.
Reasoning budgets are likely to be designed like latency and cost budgets: explicit, observable and workload-specific.
Vector Search Is Only One Retrieval Operator.
Keyword search, graph traversal, structured filters and learned routing remain important in production retrieval systems.
Observability for Probabilistic Software.
Logs tell you what happened. Traces should also preserve what the model saw, chose and rejected.