Section 02
Agents, RAG, enterprise transformation, and AI engineering.
Start here
This is a place for both: how to design and operate reliable AI systems, and how to embed them in real work so they create a durable result. Start with the question most useful to you now.
Build and operate
How to build agent systems that keep going after a failure, rather than stopping at the demo.
Redesign the work
Going beyond AI adoption to redesign the work and decisions that make up an organization.
Latest notes
Experiments and notes outside the two series, collected in reverse chronological order.
From a Dentist’s AI Stack to the Reinvention of Enterprises and Private Equity
Model child lifecycles, isolated state, and partial failure as one execution tree
What changes when retrieval becomes part of an end-to-end task?
Separate graph state from application state to assign responsibility across workers, Redis, databases, and browsers
Safely stream snapshots and events over an HTTP connection with async generators
From a convincing demo to a predictable production system
Linking user intent to side effects so a transfer tool does not run twice
Restore the screen from a snapshot, then keep its timeline complete with cursors
From prescribing every route to designing the environment and boundaries around an agent
Durable records, hot projections, and resumable snapshots survive different failures
Diagnosing bottlenecks with TTFT, TPOT, and KV cache
Separate concurrency primitives from ownership of the execution lifecycle
Serving LLMs when GPUs are scarce
Waiting, blocking SDKs, and CPU-bound computation require different execution tools
Turn time spent waiting on LLMs, databases, and search APIs into concurrency
Deciding internal open-model serving for security, data control, cost, and performance
Execution architecture boundaries — mental model, failure modes, optional deep dives
Compose bounded control flows instead of building one giant autonomous loop