FIELD NOTES
Writing.
How the systems work, where they break, and what I learned building them.
The Sandbox Wars
The tradeoffs between users, containers, sandboxes, and VMs when running untrusted agent code.
Benchmarking LLM Inference: The Metrics That Actually Matter
Latency, throughput, and the measurements that make inference benchmarks useful.
From RAG to Agentic Retrieval
From embeddings and hybrid search to retrieval agents that reflect, use tools, and try again.
From DDP to ZeRO-2 and ZeRO-3
How sharding model state saves GPU memory, and the communication costs that come with it.
Distributed Data Parallel (DDP) for Training Models
Scaling training across GPUs with replicated models, synchronized gradients, and a simple DDP implementation.
Training LLMs 101
A practical map of pre-training, fine-tuning, alignment, and optimization.
Ray for LLM Inference
Using Ray actors and resource scheduling to distribute inference across a cluster.
vLLM: LLM Inference That Doesn't Waste Your GPU
KV caches, batching, and how vLLM puts GPU memory to better use during inference.
Three Practical Ways to Detect Sensitive Data
Runnable approaches to sensitive-data detection with Presidio, GLiNER, and a hybrid of the two.
Evals: How to Evaluate Agents
Building an evaluation process around the unpredictable behavior of agents.
Why Multi-Agent Systems Matter
Patterns for dividing work among agents, coordinating their roles, and checking their results.
From Zero to Agent: ReAct, Reflection, and Planning
How tool use, reasoning, memory, and planning come together in an agent loop.
How Agents Remember: On Memory and the Art of Context Engineering
Managing context windows and deciding what an agent should carry into its next turn.
Structured Outputs in Practice: Instructor vs PydanticAI vs BAML
Comparing schemas, retries, and reliability across three structured-output frameworks.
Structured Output
Turning free-form model responses into typed data that applications can validate and use.