Dmitrii Konyrev

CTO of Argmin AI — an inference-optimization and evaluation engine for LLM-based systems (auto-optimizes model choice, prompts, and routing to cut cost 2-5x while holding quality).

• Own product + engineering end to end: architecture, roadmap, core eval/routing stack.

• Built & shipped "Judge Builder" — turns example data + expert feedback into calibrated LLM-as-a-judge evaluators (auto cold-start + human-steered calibration).

• Focus areas: LLM evaluation, guardrails, prompt/model routing, inference-cost optimization, AI-infra vendor landscape.

🔗 argminai.com/ai-quality-without-ml-team · demo youtu.be/YxccSQu_XEU · app.arcade.software/share/7Zn32dO8upImEtjQsylh

SuperAnnotate is a Series-B annotation-tech startup building end-to-end data pipelines for AI model training and validation across vision and text modalities.

• Managed an international, cross-functional team of 3 ML Engineers and 1 Product Manager, focused on productized delivery of annotation and evaluation tools.

• Launched automatic annotation tools for images, videos, and text — helping regain feature parity with competitors and significantly reduce manual labeling effort and turnaround time.

• Built semantic data curation features using embeddings (BLIP, OpenCLIP, PyTorch Metric Learning) and OpenSearch vector DB — enabling users to manage datasets more effectively and manually select similar objects for labeling or review.

• Developed LLM evaluation pipelines using DeepEval and DeepTeam, including LLM-as-a-judge scoring with integrated guardrails — automating QA workflows and improving customer confidence in label quality.

• Created agentic pipelines for pre-annotation using automatic and semi-automatic feedback loops — improving annotation accuracy and reducing human-in-the-loop load.

• Established an MLOps pipeline in collaboration with SRE and QA teams using AWS, MLflow, and Jenkins — automating quality checks and reducing deployment overhead.

• Led internal research on detecting AI-generated text and identifying impactful fine-tuning samples for non-STEM datasets — resulting in a benchmarked model published on Hugging Face, using TRL, PEFT, BitsAndBytes, Ollama, and DeepEval.

Links & resources:

🔗 https://www.superannotate.com/image-annotation-tool

🔗 https://www.superannotate.com/video-annotation

🔗 https://www.superannotate.com/text-annotation

🔗 https://www.superannotate.com/data-curation

🔗 https://huggingface.co/SuperAnnotate/ai-detector

🔗 https://www.superannotate.com/blog/superannotate-nvidia

Managed 5 risk-oriented ML engineers at Bank DOM.RF — a government-backed bank (real-estate lending, SME financing, mortgage securitization).

• Introduced bureau-based credit scoring for individuals, entrepreneurs, SMEs → data-driven loan approval.

• Launched a real-estate scoring service for B2B clients (construction/property) → automated creditworthiness, cut manual review.

• Delivered a cash-flow forecasting model for mortgage securitization → portfolio transparency, risk-based pricing.

▸ Expertise: applied ML in banking — credit scoring, risk modeling, underwriting automation in regulated environments.