Architecture of artificial intelligence systems
I design systems on large language models and write about how they are built
Notes on the architecture of artificial intelligence systems: design, engineering, operations, and the practice of building reliable solutions.
About me
more →My name is Maxim Birkin, I am an AI solutions architect. About 13 years in IT, including several years of designing LLM-based systems and taking them to production, from architecture to operations under load.
- RAG pipelines
- agent scenarios
- vector search
- streaming inference
- GPU cluster sizing
- MLOps and SRE
Articles
all →- 01
Latency engineering: analyzing 10,000 log lines in 2 minutes
How a set of deliberate levers, from parallelism to merging two model calls into one response, cut the full log-analysis time by an order of tens.
- 02
Query routing with SetFit: a few-shot classifier on CPU
Why a lightweight trained classifier stands between the code and the large model, how SetFit learns from a handful of examples, and when a query escalates to the LLM.
- 03
Cascade chains: how to separate the cause from its spreading effect
Why five affected services are most often one root cause plus four relays of its failure, and how cascade analysis narrows the hypothesis space.
- 04
No quote, no claim: how to ground a model's confidence
Anti-hallucination guards for open-ended root cause analysis: multiplying the model's self-assessment by the share of its evidence actually found in the sources.
- 05
Bayesian confidence scoring: group-max, a hard cap, and confidence bands
How to build verdict confidence from triggered detectors with deterministic code, without trusting a number from the language model, and why group-max beats the product.
- 06
A deterministic aggregator in one O(N) pass
How to remove the language model from the log-chunking stage and replace it with a statistical aggregator that computes 16 blocks of facts in one linear pass.
- 07
Inversion of responsibility: code counts the facts, the model owns the language
The central principle of a mature log-analysis system: divide responsibility between deterministic code, a trained classifier, and a large language model.