Writeups on relevant toics
Long, honest walk-throughs of the systems I build and the ideas I keep chasing, from production AI to quantum computing. If it's written up here, I actually ran it.
Computing with fragile chances, how a quantum processor really works
Read a qubit and you get a random bit; the reading itself destroys the state. It sounds like a machine made of flaws. This is the long, patient version of why it isn't, following one program from those two rules all the way down into a superconducting chip and back out as a bar chart. Grover's search worked through by hand, the transmon and its readout taken apart down to the physics, and a Braket lab you can actually run.
Enterprise AI architecture, from an LLM demo to a production system
In a production system, the language model is one component of nine. The other eight are what make it safe to operate. A capstone walk through all nine layers, organized around the failures that force each one into being and mapped to the AWS services that deliver them, from the API gateway and retrieval through validation, response, and the cross-cutting governance band.
Inside mixture of experts, how sparse models decouple size from cost
For most of the history of large models, capability and cost were the same dial: a bigger model was a more expensive model, because every parameter did work on every token. Mixture of experts broke that link. A long look at how it actually works, from the gating arithmetic to load balancing to the networking, and at where the cost really goes.
An engineering control plane for AI governance, across NIST, ISO 42001, and the EU AI Act
Three instruments now shape how organizations are expected to govern artificial intelligence, and they appear to be three separate compliance burdens. Read carefully, they describe substantially the same controls in three forms, so an engineering team can build the controls once and produce the evidence each one asks for. This is how I would structure that work.
Hardening the Model Context Protocol, a secure gateway for tool using agents
The Model Context Protocol gave agents a standard way to reach tools, and the same standardization gave attackers a predictable way to reach agents. A runnable gateway that authenticates the caller, enforces a default deny allow list, validates input, and treats tool output as untrusted data, so a confused deputy attack is blocked at the boundary instead of deleting your records.
Agentic loops, large context, and the discipline of optimization
An agentic loop grows its own context on every iteration. Every iteration appends tool calls and tool results back into the window, and on a stateless endpoint you resend that history on the next turn. This is the token math behind runaway cost, and the engineering toolkit that keeps loops predictable.
Putting a circuit breaker on your language model spending
The hardest thing about budgeting a language model call is that you do not know what it costs until after it has run. This is a small middleware you can read in one sitting that meters every call and refuses the ones you cannot afford, with per tenant budgets, rate limits, and a spend kill switch, built with the discipline of a payments authorization system.
The economics of agentic workflows, optimizing for cost and latency
A system of cooperating agents can be more capable than a single model, and it is usually more expensive to run. Whether you can run such a system at scale depends on understanding where the cost comes from, when it is worth paying, and which levers control it. It examines those economics using published figures.
Retrieval augmented generation in production, an architecture that survives contact with real data
The tutorial version of retrieval augmented generation is a four stage diagram that only covers the case where everything works. The production version is a pipeline where most of the effort goes into retrieval, evaluation, and failure modes, because a fluent answer grounded in the wrong document is worse than no answer at all. This post describes the architecture I would deploy in production.
Production architectures for AI agents, from single agents to supervisor graphs
Most agent projects fail in the gap between a working demonstration and a system you can operate. A demonstration runs once under ideal conditions; a production system has to resume after a failure, wait for a human, recover its state, stay within a budget, and leave an audit trail. This is the reference architecture that closes that gap.
Securing a language model agent the way you would secure a payments system
Prompt injection is not a minor or novelty vulnerability. It is a failure at the boundary between trusted and untrusted text, the same class of problem behind the oldest attacks in web security. This is a lab you can run that shows the attack landing, then shows three layers of defense stopping it, mapped to the frameworks security teams actually audit against.
Grover's algorithm on Amazon Braket, and what one search actually costs
A working search over two qubits that you can run for free on a simulator and, with a single flag, on real quantum hardware. The point of interest is not that it works, but the gap between the simulator and the machine, and the line on your bill.