How Local LLM Throughput Scales with Multiple Agents
Adding agents does not simply divide a local LLM’s solo token rate between them. Total output can rise even while each agent slows down.
Read articleAleksei Bukhalov
I test backend systemsunder load and failure.
SDET at MariaDB, working on distributed systems, performance, and test infrastructure. Outside work, I experiment with local LLMs and build agent tools.
Latest writing
All writingAt work
Databases, distributed systems, and performance.
I test how systems behave under load, during failures, and across distributed components. A lot of the job is building the infrastructure needed to make those tests trustworthy.
Backend and databases
Correctness, integration, failure modes, and the bugs that only appear with real data and real dependencies.
Performance and distributed systems
Load, bottlenecks, race conditions, and behavior that is easy to miss when every component is tested alone.
AWS and test infrastructure
Environments, CI/CD, observability, and tooling that make difficult failures repeatable and useful.
Projects and experiments
Selected projects
Research and engineering tools.
Experiments with explicit limits, tools for practical work, and infrastructure for testing it.
01compact_llm_languageCompleted research · local LLMsControlled experiments in hidden-state recurrence for Qwen3.5-2B. The tested recipes produced no replicated reasoning gain; the methods, controls, and negative results are documented.02requirements-minerAgent tooling · requirementsRequirements tooling that runs isolated expert reviews, weighs evidence, and preserves disagreement. One shared skill across multiple agent harnesses.03serbian-dubPython · end-to-end media pipelineA Russian-to-Serbian video dubbing pipeline: transcription, speaker-aware translation, voice synthesis, timing, mixing, and automated QA.
More projects on GitHub Small iOS experiments
Small product experiments built to explore narrow ideas and learn the full path from prototype to App Review.
View all experiments