Notebook

Writing

Engineering notes on AI execution systems, platform architecture, and operational reliability.

These are notes from building systems: execution runtimes, policy-aware workflows, replayable AI pipelines, model boundaries, and platform reliability.

Back to homepage writing teaser

Featured theme

AI Execution Systems

This series explores how AI workflows behave when models, tools, policies, retries, validation, replay, and operators are treated as first-class system concerns.

Published / Available writing

LLM Foundationspublished11 min read

How Large Language Models Actually Work: The Foundations

A ground-up tour of how large language models actually work: tokens, embeddings, the Transformer and attention, parameters, training versus inference, open versus closed models, and the classical ML it is all built on.

LLMs
AI Engineering
Transformers
Machine Learning
Foundations

Read article

AI Execution Systemspublished8 min read

Why AI Systems Need Execution Semantics

Most AI demos describe what the model should do. Production systems need to describe what actually ran, why it ran, what was allowed, and how it can be replayed.

AI Systems
Orchestration
Reliability
Platform Engineering

Read article

AI execution systemspublished

Designing an LLM orchestrator for production systems

How production orchestration differs from chaining prompts: explicit state, scheduling, tool contracts, and failure semantics when every step can stall or lie.

orchestration
execution semantics
tooling

Read article

RAG and retrievalpublished

Why most RAG systems fail in production

Failure modes that benchmarks skip: drift, access control, stale corpora, evaluation gaps, and the gap between “retrieved” and “usable under load.”

rag
retrieval
production

Read article

AI platform architecturepublished

End-to-end flow of an AI request in production systems

Tracing a request through routing, policy, retrieval, inference, and side effects—where latency and partial failure actually show up.

platform
policy
runtime

Read article

ML fundamentalspublished

Building a ternary sentiment classifier with modern NLP

End-to-end NLP pipeline and evaluation, not a toy binary task; useful as a sample of rigorous problem framing.

nlp
evaluation

Read article

MLOpspublished

Deploying a production-ready model: Titanic survival

Deployment hygiene and packaging as release discipline: useful as a checklist for moving models out of notebooks.

deployment
mlops

Read article

Planned essays

AI execution systems

AI Execution Systems

How AI workflows behave when models, tools, policy, retries, validation, replay, and operators are treated as first-class runtime concerns.

AI Execution Systemsplanned11 min

Replayability in AI Workflows

Design patterns for reruns, replay children, and reproducible debugging under production drift.

replay
debugging

AI Execution Systemsplanned13 min

Deterministic Orchestration vs Agent Loops

Tradeoffs between bounded execution graphs and open-ended model loops in enterprise systems.

orchestration
agent loops

AI Execution Systemsplanned10 min

Policy-Aware AI Execution

How to separate policy evaluation from model reasoning and still keep operator velocity.

policy
governance

AI Execution Systemsplanned12 min

Building a Bounded Model Runtime

Runtime boundaries, structured outputs, and guardrails for model calls in multi-step workflows.

model runtime
guardrails

AI Execution Systemsplanned11 min

Designing AI Systems Operators Can Debug

Trace shape, event contracts, and UX patterns that make incidents diagnosable.

observability
operator ux

AI Execution Systemsplanned14 min

Tool Runtime Design for Agentic Systems

Typed contracts, side-effect controls, and failure envelopes for tool invocation at scale.

tool runtime
safety

AI Execution Systemsplanned12 min

Post-Execution Learning Without Runtime Self-Mutation

How to improve systems with offline insights while keeping runtime behavior deterministic.

post-execution
learning loop

AI platform architecture