esmirnov.aiAI projects I have worked on.
AI tool tests
ToolBoundary finds differences between parser output and the values that an AI tool handler or client receives. It uses stub handlers to examine tool inputs, results, and errors through SDK paths.
The Python package supports MCP Python SDK v1. The Node.js package supports MCP TypeScript SDK v1 and AI SDK WorkflowAgent v2.
AI research workspace
NematicLab takes a research question through literature research, scientific calculations, and manuscript preparation. A researcher can add papers, equations, or code. AI helps find related work and use symbolic or numerical tools. It can draft Methods and Results from saved sources and calculations.
Sources, plots, calculation records, and manuscript versions stay in one study. Researchers examine the science, revise the draft, and export the article, PDF, figures, and supporting files. The site shows a preview of the workbench.
Assessment workflows
This workspace takes an assessment from assignment and submission through a private AI draft, teacher review, and released feedback. I made the scoring and review flow, with student and teacher interfaces.
An AI companion uses class and student records to prepare more work. Revision tests compare student changes with the agreed task. Teachers approve recommendations before the workspace assigns or releases work.
Agent workflows
A stateful agent works in web and mobile trading products. It uses live market, news, and trading data. It calls platform tools and does scheduled monitoring and longer analyses. Checkpoints and admin controls manage the agent state and behavior.
Model training and orchestration
I made a trading chatbot with deterministic rules, XLM-R classification, and Qwen. I fine-tuned and distilled the models. The routing and inference paths support requests with low latency.
Conversational workflows
A LangGraph chatbot matches a visitor's needs to available solutions. It uses retrieval and a guided conversation to record qualified leads.
English-writing practice · Closed
I made an English writing assessment platform before the ChatGPT era. It had a Python backend, AI models, and React frontends. BERT, spaCy, and NLTK assessed task achievement, coherence, cohesion, and vocabulary range.
I trained XGBoost regressors on examiner-scored essays to predict writing scores. The platform supported IELTS practice through an API, learner applications, and delivery for education businesses. It gave feedback in web reports and PDF statements.