Production AI systems
About Slava Dubrov
Production AI systems, evaluation, security, and the evidence loops that improve them.
I’m Slava Dubrov—formally, Dr. Viacheslav Dubrov*—a Staff AI Systems Engineer working across software engineering, agent systems, evaluation, retrieval, security, and production infrastructure.
These days I work at octonomy AI on production AI and software systems. Before that, at HubSpot, I led work on the Context Layer for retrieval, grounding, and memory infrastructure, then worked on Agent Execution across LLM deployment, fine-tuning, evaluation, safety, and runtime. The connecting thread is what happens after a promising model output meets a real system.
In my current writing and open-source work, I explore how agents can help improve other AI systems. An LLM agent investigates a failure, proposes a change within defined limits, and uses test results to decide what to try next. I call this recursive harness engineering, one harness improving another; my series on it began as Meta-Engineering AI Systems. People set the objectives, allowed changes, budget, and criteria for adopting a result. The process is closed-loop because each result informs the next experiment. This is recursive improvement, not self-improvement: the agent never edits its own harness or the checks that judge it.
The work behind the point of view
My work has moved from ranking and fraud systems to retrieval infrastructure and production agents.
Selected experience
- HubSpot — Context Layer and Agent Execution. I owned Staff-level architecture and technical leadership across retrieval, grounding, memory, LLM deployment, fine-tuning, evaluation, safety, and runtime.
- Wayfair — fraud, scam detection, and embeddings. I built and led systems associated with about $4M in annual savings.
- OLX — recommendation and search ranking. I worked on ranking and recommendation systems at marketplace scale.
Public evidence
- Labs collects reproducible artifacts that can be run locally.
- Recursive Harness Engineering develops the improvement process through runnable experiments, starting with an agent that proposes and tests changes to a memory tool.
- Edge of Context holds the long-form engineering record, including the six-part Engineering the Agentic Stack series.
Research and speaking
I hold a Kandidat of Technical Sciences degree (Russian doctoral degree, PhD-equivalent) from Southern Federal University for research in AI diagnostics, with peer-reviewed papers and patents. I also spent a doctoral research semester in Mechatronics at TU Ilmenau on a DAAD Lomonosov scholarship.
At World Agentic AI Summit Berlin 2026, I presented “Engineering the Agentic Stack”: a production architecture covering the cognitive engine, memory, tools, security, and the runtime around them.
Engineering principles
- Make behavior observable before making it autonomous.
- Prefer environment state and outcomes over plausible final answers.
- Make each added harness component earn its cost through ablation.
- Keep evaluators, security policy, and promotion authority independent from the optimizer.
- Preserve failed experiments as evidence.
- Use the cheapest repair layer that solves the real failure.
Capabilities
| System capability | Working areas |
|---|---|
| Agent systems | reasoning loops, memory, tool interfaces, runtimes, harnesses, observability |
| Evaluation and assurance | trace-based evaluation, regression suites, guardrails, permissions, security gates |
| Retrieval and context | semantic and hybrid retrieval, reranking, grounding, context compression |
| Model engineering | fine-tuning, LoRA/QLoRA, serving, quantization, structured generation |
| Production infrastructure | AWS, GCP, Kubernetes, batch, streaming, and real-time ML systems |
Elsewhere
Professional background is on LinkedIn. Source and reproducible artifacts are on GitHub. New essays also appear on Substack.
* Kandidat of Technical Sciences (Russian doctoral degree, PhD-equivalent), awarded by Southern Federal University, Rostov-on-Don, 2015.