▦ Engineering Vault

The whole agent stack, bottom-up — tell real capability from hype in pitches and reviews.

Why one AI system feels far smarter than another on the identical model — selection, retrieval & assembly of context.

How AI systems are measured and made to improve — offline & online eval, golden sets, LLM-as-judge, and the production eval loop.

Review AI systems with the rigor of production distributed systems — reliability, scalability, observability, cost, governance & failure modes.

The AI attack surface — prompt injection, jailbreaks, tool abuse, data leakage & multi-tenant risk — with threat models, detection & mitigation.

The infrastructure powering AI products — embeddings, vector DBs, rerankers, model gateways, inference & caching — with scaling & cost trade-offs.

A repeatable method to design production agents for real domains — responsibilities, tools, context, memory, runtime, eval, security, approval & ROI.

Advanced AI systems judged on engineering merit — multi-agent & swarms, fine-tuning, distillation, reasoning models & long-horizon autonomy: when, and when not.

The design decisions a senior interview probes — each infographic-first with a quiz and a rehearse-out-loud drill.

Foundation → Staff: the membership-test asymmetry, sizing math, failure modes, and variants at scale.

Why distributed systems are hard, and the primitives that tame failure and time.

How databases actually store and find data — and the read/write/space trade-offs underneath.

What isolation levels really promise, and how MVCC and SSI deliver them.

The log as a primitive, and how exactly-once and event-time processing are really achieved.

A repeatable method for system-design rounds, plus the building blocks they lean on.