METAHEURISTICBook a call
[Curriculum vitae]

Alexander Myasoedov

Summary

I build systems where a language model plans, calls tools, and acts - and I make those systems measurable, safe and cheap enough to point at production traffic.

The work is closer to distributed systems engineering than to modelling: control flow with real budgets, tool contracts that survive an unreliable caller, retrieval that respects your permission model, an eval harness that turns prompt changes into diffs with numbers attached, and traces that make a 3am diagnosis possible.

I also maintain the open-source tooling other teams use to find these failures - including an LLM red-teaming kit with around two thousand stars - and I write about the operational side of this work in public, weekly.

Available for contract, fractional and embedded engagements. Code, prompts, evals and IP are yours from day one.

Stack
Agentic systems
Agent control flow Tool contracts & MCP Loop guards & budgets Human-in-the-loop approvals Durable execution (Temporal) LangGraph Multi-agent orchestration
Model layer
Eval harnesses & golden sets LLM-as-judge calibration Prompt versioning Model routing & fallback Prompt caching Cost & latency budgeting Fine-tuning decisions vLLM / open-weight serving
Retrieval
Permission-aware RAG Hybrid search (BM25 + dense) Reranking Layout-aware PDF parsing Structure-aware chunking pgvector Freshness & incremental indexing
Security & governance
LLM red-teaming Prompt injection defence Tool-scope least privilege Output validation PII handling Audit logging On-premise & sovereign deploy
Engineering
Python Go Elixir Haskell FastAPI PostgreSQL Kafka Docker Kubernetes OpenTelemetry AWS / GCP
Experience
Founder & principal engineer · Metaheuristic
Present

Applied AI partner. Takes LLM prototypes into production for teams that need the system to survive real users, real documents and a real bill.

  • Ships agentic systems end to end: control flow, tool integrations, retrieval, evals, guardrails, traces and handover.
  • Cut inference spend on one production system from roughly $4,120 to $1,180 a month through caching and model routing, at the same quality and latency.
  • Runs fixed-scope Discovery Sprints that take teams from demo to production in two to ten weeks, with evals and a security review gating every release.
  • Delivers on-premise and sovereign deployments for teams that cannot send data to a third-party API.
  • Client keeps 100% of code, prompts, evals and IP - no runtime licence, no platform lock-in.
Open-source author & maintainer · github.com/msoedov
2012 - present

57 public repositories, roughly 4,600 stars, across LLM security, LLMOps, retrieval and developer tooling.

  • agentic_security (~2.0k stars) - agentic LLM vulnerability scanner and red-teaming kit, used by other teams to test their own agents for jailbreaks, prompt injection and data leakage.
  • langcorn (938 stars) - serves LangChain LLM apps and agents through FastAPI; the LLMOps layer between a notebook and a callable API.
  • validex (144 stars) - structured extraction from unstructured sources, the unglamorous half of every RAG project.
  • justshowmediff (69 stars) - browser-based review of AI-generated diffs, built for Claude Code, Codex and headless agent workflows.
  • vector_lake (59 stars) - S3-backed vector database for LLM agents and RAG, for teams who want retrieval without a managed vendor in the critical path.
Writing · metaheuristic.co/blog
Ongoing

Essays on the operational side of production AI - agent control flow, evals, review markets, retrieval, and keeping a human in the loop without slowing the loop down.

Need an agentic AI engineer this quarter?

Contract, fractional, or embedded with your team. Start with a 20-minute call - you leave it with an honest read on whether your use case is buildable, whatever happens next.